1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242 3243 3244 3245 3246 3247 3248 3249 3250 3251 3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371 3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516 3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549 3550 3551 3552 3553 3554 3555 3556 3557 3558 3559 3560 3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593 3594 3595 3596 3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617 3618 3619 3620 3621 3622 3623 3624 3625 3626 3627 3628 3629 3630 3631 3632 3633 3634 3635 3636 3637 3638 3639 3640 3641 3642 3643
|
// earth.c
//
// This code is derived from code in the Rational Fortran file dmarss.r which is
// part of the R and S mda package by Hastie and Tibshirani.
// Comments containing "TODO" mark known issues
//
// See the R earth documentation for descriptions of the principal data structures.
// See also www.milbo.users.sonic.net. This code uses a subset of C99.
//
// Stephen Milborrow Feb 2007 Petaluma
//
// References:
//
// HastieTibs: Trevor Hastie and Robert Tibshirani
// S library mda version 0.3.2 dmarss.r Ratfor code
// Modifications for R by Kurt Hornik, Friedrich Leisch, Brian Ripley
//
// FriedmanMars: Multivariate Adaptive Regression Splines (with discussion)
// Annals of Statistics 19/1, 1--141, 1991
//
// FriedmanFastMars: Friedman "Fast MARS"
// Dep. of Stats. Stanford, Tech Report 110, May 1993
//
// Miller: Alan Miller (2nd ed. 2002) Subset Selection in Regression
//
//-----------------------------------------------------------------------------
// This program is free software; you can redistribute it and/or modify
// it under the terms of the GNU General Public License as published by
// the Free Software Foundation; either version 2 of the License, or
// (at your option) any later version.
//
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License for more details.
//
// A copy of the GNU General Public License is available at
// http://www.r-project.org/Licenses
//
//-----------------------------------------------------------------------------
#if !STANDALONE
#define USING_R 1
#endif // STANDALONE
#include <stdlib.h>
#include <stdio.h>
#include <stdarg.h>
#include <string.h>
#include <float.h>
#include <math.h>
#if _MSC_VER // microsoft
#include <crtdbg.h> // microsoft malloc debugging library
#define _C_ "C"
// disable warning: 'vsprintf': This function or variable may be unsafe
#pragma warning(disable: 4996)
#else
#define _C_
#ifndef bool
typedef int bool;
#define false 0
#define true 1
#endif
#endif
#if USING_R // R with gcc
#include "R.h"
#include "Rinternals.h" // needed for Allowed function handling
#include "allowed.h"
#include "R_ext/Rdynload.h"
#define printf Rprintf
#define FINITE(x) R_FINITE(x)
#else
#define warning printf
void error(const char* args, ...);
#if _MSC_VER // microsoft
#define ISNAN(x) _isnan(x)
#define FINITE(x) _finite(x)
#else
#define ISNAN(x) isnan(x)
#define FINITE(x) finite(x)
#endif
#endif
#include "earth.h"
#ifdef MATLAB
#include "mex.h" // for printf
#endif
// linpack functions (dqrdc2 is a modified form of dqrdc used by R)
extern _C_ int dqrdc2_(double* x, int* ldx, int* n, int* p,
double* tol, int* rank,
double* qraux, int* pivot, double* work);
extern _C_ int dtrsl_(double* t, int* ldt, int* n, double* b, int* job, int* info);
extern _C_ int dqrsl_(double* x, int* ldx, int* n, int* k,
double* qraux, double* y,
double* qy, double* qty, double* b,
double* rsd, double* xb, int* job, int* info);
extern _C_ int daxpy_(const int* n, const double* alpha,
const double* dx, const int* incx,
double* dy, const int* incy);
extern _C_ double ddot_(const int* n,
const double* dx, const int* incx,
const double* dy, const int* incy);
#define ASSERT(x) \
if(!(x)) error("internal assertion failed in file %s line %d: %s\n", \
__FILE__, __LINE__, #x)
#define sq(x) ((x) * (x))
#ifndef max
#define max(a,b) (((a) > (b)) ? (a) : (b))
#endif
#ifndef min
#define min(a,b) (((a) < (b)) ? (a) : (b))
#endif
#define INLINE inline
#define USE_BLAS 1 // 1 is faster (tested on Windows XP Pentium with R BLAS)
// also, need USE_BLAS to use bxOrthCenteredT
#define FAST_MARS 1 // 1 to use techniques in FriedmanFastMars (see refs)
#define WEIGHTS 1 // 1 if case weights are supported
#if STANDALONE
#define IOFFSET 0
#else
#define IOFFSET 1 // printfs only: 1 to convert 0-based indices to 1-based in printfs
// use 1 for R style indices in messages to the user
// use 0 for C style indices in messages to the user
#endif
static const char* VERSION = "version 4.6.3"; // change if you modify this file!
static const double MIN_GRSQ = -10.0;
static const double QR_TOL = 1e-8; // same as R lm
static const double MIN_BX_SOS = .01;
static const double ALMOST_ZERO = 1e-10;
static const int ONE = 1; // parameter for BLAS routines
#ifdef __INTEL_COMPILER
static const double POS_INF = __builtin_inff();
#elif _MSC_VER // microsoft compiler
static const double POS_INF = _HUGE;
#else // assume gcc
static const double POS_INF = (1.0 / 0.0);
#endif
static const int MAX_DEGREE = 100;
// Poor man's array indexing -- not pretty, but pretty useful.
//
// Note that we use column major ordering. C programs usually use row major
// ordering but we don't here because the functions in this file are called
// by R and call Fortran routines which use column major ordering.
//
// Note that nCases is size_t (not int), allowing array indices bigger than 2GB.
// We don't expect nCases itself to be that big, but it can be used in
// expressions that evaluate to more than 2GB.
#define Dirs_(iTerm,iPred) Dirs[(iTerm) + (iPred)*(nMaxTerms)]
#define Cuts_(iTerm,iPred) Cuts[(iTerm) + (iPred)*(nMaxTerms)]
#define bx_(i,iTerm) bx [(i) + (iTerm)*(nCases)]
#define bxUsed_(i,iTerm) bxUsed [(i) + (iTerm)*(nCases)]
#define bxOrth_(i,iTerm) bxOrth [(i) + (iTerm)*(nCases)]
#define bxOrthCenteredT_(iTerm,i) bxOrthCenteredT[(iTerm) + (i)*(nMaxTerms)]
#define x_(i,iPred) x [(i) + (iPred)*(nCases)]
#define xOrder_(i,iPred) xOrder [(i) + (iPred)*(nCases)]
#define y_(i,iResp) y [(i) + (iResp)*(nCases)]
#define Residuals_(i,iResp) Residuals [(i) + (iResp)*(nCases)]
#define ycboSum_(iTerm,iResp) ycboSum [(iTerm) + (iResp)*(nMaxTerms)]
#define Betas_(iTerm,iResp) Betas [(iTerm) + (iResp)*(nUsedCols)]
// Global copies of some input parameters. These stay constant for the entire MARS fit.
static double TraceGlobal; // copy of Trace parameter
static int nMinSpanGlobal; // copy of nMinSpan parameter
static int nEndSpanGlobal; // copy of nEndSpan parameter
static double AdjustEndSpanGlobal; // copy of AdjustEndSpan parameter
static void FreeBetaCache(void);
static char* sFormatMemSize(const size_t MemSize, const bool Align);
//-----------------------------------------------------------------------------
// DON'T USE free, malloc, and calloc in this file.
// Use free1, malloc1, and calloc1 instead.
//
// malloc and its friends are redefined (a) so under microsoft C using
// crtdbg.h we can easily track alloc errors and (b) so FreeR() doesn't
// re-free any freed blocks and (c) so out of memory conditions are
// immediately detected.
// free1 is a macro so we can zero p (so we know it is released if FreeR is invoked)
#define free1(p) \
{ \
if(p) \
free(p); \
p = NULL; \
}
static void* malloc1(size_t size, const char* args, ...)
{
void* p = malloc(size);
if(!p || TraceGlobal == 1.5) {
if(args == NULL)
printf("malloc %s\n", sFormatMemSize(size, true));
else {
char s[1000];
va_list va;
va_start(va, args);
vsprintf(s, args, va);
va_end(va);
printf("malloc %s: %s\n", sFormatMemSize(size, true), s);
}
}
if(!p)
error("Out of memory (could not allocate %s)", sFormatMemSize(size, false));
return p;
}
static void* calloc1(size_t num, size_t size, const char* args, ...)
{
void* p = calloc(num, size);
if(!p || TraceGlobal == 1.5) {
if(args == NULL)
printf("calloc %s\n", sFormatMemSize(size, true));
else {
char s[1000];
va_list va;
va_start(va, args);
vsprintf(s, args, va);
va_end(va);
printf("calloc %s: %s\n", sFormatMemSize(size, true), s);
}
}
if(!p)
error("Out of memory (could not allocate %s)", sFormatMemSize(size, false));
return p;
}
// After calling this, on program termination we will get a report if there are
// writes outside the borders of allocated blocks or if there are non-freed blocks.
#if _MSC_VER && _DEBUG // microsoft C and debugging enabled?
static void InitMallocTracking(void)
{
_CrtSetReportMode(_CRT_WARN, _CRTDBG_MODE_WNDW);
_CrtSetReportMode(_CRT_WARN, _CRTDBG_MODE_FILE);
_CrtSetReportFile(_CRT_WARN, _CRTDBG_FILE_STDOUT);
int Flag = _CrtSetDbgFlag(_CRTDBG_REPORT_FLAG);
Flag |= (_CRTDBG_ALLOC_MEM_DF|
// following commented out because it makes earth run very slowly
// _CRTDBG_CHECK_ALWAYS_DF| // call _CrtCheckMemory at every alloc and free
_CRTDBG_LEAK_CHECK_DF|
_CRTDBG_DELAY_FREE_MEM_DF);
_CrtSetDbgFlag(Flag);
}
#endif
//-----------------------------------------------------------------------------
// These are malloced blocks. They unfortunately have to be declared globally so
// under R if the user interrupts we can free them using on.exit(.C("FreeR"))
static int* xOrder; // local to FindTerm
static bool* WorkingSet; // local to FindTerm and EvalSubsets
static double* xbx; // local to FindTerm
static double* CovSx; // local to FindTerm
static double* CovCol; // local to FindTerm
static double* ycboSum; // local to FindTerm (used to be called CovSy)
static double* bxOrth; // local to ForwardPass
static double* yMean; // local to ForwardPass
// Transposed and mean centered copy of bxOrth, for fast update in FindKnot.
// It's faster because there is better data locality as iTerm increases, so
// better L1 cache use. This is used only if USE_BLAS is true.
static double* bxOrthCenteredT; // local to ForwardPass
static double* bxOrthMean; // local to ForwardPass
static int* nDegree; // local to Earth or ForwardPassR
static int* nUses; // local to Earth or ForwardPassR
#if USING_R
static int* iDirs; // local to ForwardPassR
static bool* BoolFullSet; // local to ForwardPassR
#endif
#if FAST_MARS
static void FreeQ(void);
#endif
#if USING_R
void FreeR(void) // for use by R
{
free1(WorkingSet);
free1(CovSx);
free1(CovCol);
free1(ycboSum);
free1(xOrder);
free1(bxOrthMean);
free1(bxOrthCenteredT);
free1(bxOrth);
free1(yMean);
free1(BoolFullSet);
free1(iDirs);
free1(nUses);
free1(nDegree);
FreeBetaCache();
#if FAST_MARS
FreeQ();
#endif
}
#endif
//-----------------------------------------------------------------------------
static char* sFormatMemSize(const size_t MemSize, const bool Align)
{
static char s[100];
double Size = (double)MemSize;
if(Size >= 1e9)
sprintf(s, Align? "%6.3f GB": "%.3g GB", Size / ((size_t)1 << 30));
else if(Size >= 1e6)
sprintf(s, Align? "%6.0f MB": "%.3g MB", Size / ((size_t)1 << 20));
else if(Size >= 1e3)
sprintf(s, Align? "%6.0f kB": "%.3g kB", Size / ((size_t)1 << 10));
else
sprintf(s, Align? "%6.0f B": "%g B", Size);
return s;
}
//-----------------------------------------------------------------------------
static void tprintf(const int trace, const char *args, ...) // printf with trace check
{
if(TraceGlobal >= trace) {
char s[1000];
va_list va;
va_start(va, args);
vsprintf(s, args, va);
va_end(va);
printf("%s", s);
}
}
//-----------------------------------------------------------------------------
// Gets called periodically to service the R framework.
// Will not return if the user interrupts.
#if USING_R
static INLINE void ServiceR(void)
{
R_FlushConsole();
R_CheckUserInterrupt(); // may never return
}
#endif
//-----------------------------------------------------------------------------
#if FAST_MARS
typedef struct tQueue {
int iParent; // parent term
double RssDelta;
int nTermsForRssDelta; // number of terms when RssDelta was calculated
double AgedRank;
} tQueue;
static tQueue* Q; // indexed on iTerm (this Q is used for queue updates)
static tQueue* SortedQ; // indexed on iParent rank (this Q is used to get next iParent)
static int nQMax; // number of elements in Q
static void InitQ(const int nMaxTerms)
{
nQMax = 0;
Q = (tQueue*)malloc1(nMaxTerms * sizeof(tQueue),
"Q\t\t\tnMaxTerms %d sizeof(tQueue) %d",
nMaxTerms, sizeof(tQueue));
SortedQ = (tQueue*)malloc1(nMaxTerms * sizeof(tQueue),
"SortedQ\t\tnMaxTerms %d sizeof(tQueue) %d",
nMaxTerms, sizeof(tQueue));
for(int iTerm = 0; iTerm < nMaxTerms; iTerm++) {
Q[iTerm].iParent = iTerm;
Q[iTerm].nTermsForRssDelta = -99; // not strictly needed, nice for debugging
Q[iTerm].RssDelta = -1;
Q[iTerm].AgedRank = -1;
}
}
static void FreeQ(void)
{
free1(Q);
free1(SortedQ);
}
static void PrintSortedQ(int nFastK) // for debugging
{
printf("\n\nSortedQ QIndex Parent nTermsForRssDelta AgedRank RssDelta\n");
for(int i = 0; i < nQMax; i++) {
printf(" %3d %3d %15d %5.1f %g\n",
i+IOFFSET,
SortedQ[i].iParent+IOFFSET,
SortedQ[i].nTermsForRssDelta+IOFFSET,
SortedQ[i].AgedRank,
SortedQ[i].RssDelta);
if(i == nFastK-1)
printf("FastK %d ----------------------------------------------------\n",
nFastK);
}
}
// Sort so highest RssDeltas are at low indices.
// Secondary sort key is iParent. Not strictly needed, but removes
// possible differences in qsort implementations (which "sort"
// identical keys unpredictably).
static int CompareQ(const void* p1, const void* p2) // for qsort
{
double Diff = ((const tQueue*)p2)->RssDelta - ((const tQueue*)p1)->RssDelta;
if(Diff < 0)
return -1;
else if(Diff > 0)
return 1;
// Diff is 0, so sort now on iParent
int iDiff = ((const tQueue*)p1)->iParent - ((const tQueue*)p2)->iParent;
if(iDiff < 0)
return -1;
else if(iDiff > 0)
return 1;
return 0;
}
// Sort so lowest AgedRanks are at low indices.
// If AgedRanks are the same then sort on RssDelta and iParent.
static int CompareAgedQ(const void* p1, const void* p2) // for qsort
{
double Diff = ((const tQueue*)p1)->AgedRank - ((const tQueue*)p2)->AgedRank;
if(Diff < 0)
return -1;
else if(Diff > 0)
return 1;
// Diff is 0, so sort now on RssDelta
Diff = ((const tQueue*)p2)->RssDelta - ((const tQueue*)p1)->RssDelta;
if(Diff < 0)
return -1;
else if(Diff > 0)
return 1;
// Diff is still 0, so sort now on iParent
int iDiff = ((const tQueue*)p1)->iParent - ((const tQueue*)p2)->iParent;
if(iDiff < 0)
return -1;
else if(iDiff > 0)
return 1;
return 0;
}
static void AddTermToQ(
const int iTerm, // in
const int nTerms, // in
const double RssDelta, // in
const bool Sort, // in
const int nMaxTerms, // in
const double FastBeta) // in: ageing Coef, 0 is no ageing, FastMARS recommends 1
{
ASSERT(iTerm < nMaxTerms);
ASSERT(nQMax < nMaxTerms);
Q[nQMax].nTermsForRssDelta = nTerms;
Q[nQMax].RssDelta = max(Q[iTerm].RssDelta, RssDelta);
nQMax++;
if(Sort) {
memcpy(SortedQ, Q, nQMax * sizeof(tQueue));
qsort(SortedQ, nQMax, sizeof(tQueue), CompareQ); // sort on RssDelta
if(FastBeta > 0) {
for(int iRank = 0; iRank < nQMax; iRank++)
SortedQ[iRank].AgedRank =
iRank + FastBeta * (nTerms - SortedQ[iRank].nTermsForRssDelta);
qsort(SortedQ, nQMax, sizeof(tQueue), CompareAgedQ); // sort on aged rank
}
}
}
static void UpdateRssDeltaInQ(const int iParent, const int nTermsForRssDelta,
const double RssDelta)
{
ASSERT(iParent == Q[iParent].iParent);
ASSERT(iParent < nQMax);
Q[iParent].nTermsForRssDelta = nTermsForRssDelta;
Q[iParent].RssDelta = RssDelta;
}
static int GetNextParent( // returns -1 if no more parents
const bool InitFlag, // use true to init, thereafter false
const int nFastK)
{
static int iQ; // index into sorted queue
int iParent = -1;
if(InitFlag) {
if(TraceGlobal == 6)
printf("\n|Considering parents ");
iQ = 0;
} else {
if(iQ < min(nQMax, nFastK)) {
iParent = SortedQ[iQ].iParent;
iQ++;
}
if(TraceGlobal == 6 && iParent >= 0)
printf("%d [%g] ", iParent+IOFFSET, SortedQ[iQ].RssDelta);
}
return iParent;
}
#endif // FAST_MARS
//-----------------------------------------------------------------------------
// This helps reduce the effects of numerical err, mostly when testing.
static INLINE double MaybeZero(double x)
{
return x < ALMOST_ZERO? 0. : x;
}
//-----------------------------------------------------------------------------
// GetOrder() gets the sort indices of vector x, so
// x[sorted[i]] <= x[sorted[i+1]]. Ties may be reordered. The returned
// indices are 0 based (as in C not as in R).
//
// This function is similar to the R library function rsort_with_index(),
// but is defined here to minimize R dependencies.
// Informal tests show that this is faster than rsort_with_index().
static const double* pxGlobal; // needed because of the way qsort works
static int Compare(const void* p1, const void* p2) // for qsort
{
const int i1 = *(const int*)p1;
const int i2 = *(const int*)p2;
const double Diff = pxGlobal[i1] - pxGlobal[i2];
if(Diff < 0)
return -1;
else if(Diff > 0)
return 1;
else
return 0;
}
static void GetOrder(
int sorted[], // out: vec with nx elements
const double x[], // in: x is a vec with nx elems
const int nx) // in: number of elems in x
{
for(int i = 0; i < nx; i++)
sorted[i] = i;
pxGlobal = x;
qsort(sorted, nx, sizeof(int), Compare);
}
//-----------------------------------------------------------------------------
// Get order indices for an x array of dimensions nRows x nCols.
//
// Returns an nRows x nCols integer array of indices, where each column
// corresponds to a column of x. See GetOrder() for ordering details.
//
// Caller must free the returned array.
static int* GetArrayOrder(
const double x[], // in
const size_t nRows, // in
const int nCols) // in
{
int* xOrder = (int*)malloc1(nRows * nCols * sizeof(int),
"xOrder\t\tnRows %d nCols %d sizeof(int) %d",
nRows, nCols, sizeof(int));
// following can be quite slow if nRows is big (requires qsort for each colun)
for(int iCol = 0; iCol < nCols; iCol++) {
GetOrder(xOrder + iCol*nRows, x + iCol*nRows, (int)nRows);
#if USING_R
if(nRows > (int)1e4)
ServiceR();
#endif
}
return xOrder;
}
//-----------------------------------------------------------------------------
// return the number of TRUEs in the boolean vector UsedCols
static int GetNbrUsedCols(const bool UsedCols[], const int nLen)
{
int nTrue = 0;
for(int iCol = 0; iCol < nLen; iCol++)
if(UsedCols[iCol])
nTrue++;
return nTrue;
}
//-----------------------------------------------------------------------------
// Copy used columns in x to *pxUsed and return the number of used columns
// UsedCols[i] is true for each each used column index in x
// Caller must free *pxUsed
static int CopyUsedCols(
double** pxUsed, // out: caller must free
const double x[], // in: nCases x nCols
const size_t nCases, // in
const int nCols, // in
const bool UsedCols[]) // in
{
const int nUsedCols = GetNbrUsedCols(UsedCols, nCols);
double* xUsed = (double*)malloc1(nCases * nUsedCols * sizeof(double),
"xUsed\t\t\tnCases %d nUsedCols %d sizeof(double) %d",
(int)nCases, nUsedCols, sizeof(double));
int iUsed = 0;
for(int iCol = 0; iCol < nCols; iCol++)
if(UsedCols[iCol]) {
memcpy(xUsed + iUsed * nCases,
x + iCol * nCases, nCases * sizeof(double));
iUsed++;
}
*pxUsed = xUsed;
return nUsedCols;
}
//-----------------------------------------------------------------------------
// Print a summary of the model, for debug tracing
#if STANDALONE
static void PrintSummary(
const int nMaxTerms, // in
const int nTerms, // in: number of cols in bx, some may be unused
const int nPreds, // in: number of predictors
const int nResp, // in: number of cols in y
const bool UsedCols[], // in: specifies used columns in bx
const int Dirs[], // in
const double Cuts[], // in
const double Betas[], // in: if NULL will print zeros
const int nDegree[]) // in: degree of each term, degree of intercept is 0
{
printf(" nFacs Beta\n");
const int nUsedCols = GetNbrUsedCols(UsedCols, nTerms);
int iUsed = -1;
for(int iTerm = 0; iTerm < nTerms; iTerm++) {
if(UsedCols[iTerm]) {
iUsed++;
printf("%2.2d %2d ", iTerm, nDegree[iTerm]);
for(int iResp = 0; iResp < nResp; iResp++)
printf("%9.3g ", (Betas? Betas_(iUsed, iResp): 0));
printf("| ");
}
else {
printf("%2.2d -- ", iTerm);
for(int iResp = 0; iResp < nResp; iResp++)
printf("%9s ", "--");
printf("| ");
}
int iPred;
for(iPred = 0; iPred < nPreds; iPred++)
if(Dirs_(iTerm,iPred) == 0)
printf(" . ");
else
printf("%2d ", Dirs_(iTerm,iPred));
printf("|");
for(iPred = 0; iPred < nPreds; iPred++)
if(Dirs_(iTerm,iPred) == 0)
printf(" . ");
else if(Dirs_(iTerm,iPred) == 2)
printf(" linear ");
else
printf("%8.3g ", Cuts_(iTerm,iPred));
printf("\n");
}
printf("\n");
}
#endif // STANDALONE
//-----------------------------------------------------------------------------
// Set Diags to the diagonal values of inverse(X'X),
// where X is referenced via the matrix R, from a previous call to dqrdc2
// with (in practice) bx. The net result is that Diags is the diagonal
// values of inverse(bx'bx). We assume that R is created from a full rank X.
//
// TODO This could be simplified
static void CalcDiags(
double Diags[], // out: nCols x 1
const double R[], // in: nCases x nCols, QR from prev call to dqrdc2
const size_t nCases, // in
const int nCols) // in
{
#define R_(i,j) R [(i) + (j) * nCases]
#define R1_(i,j) R1[(i) + (const size_t)(j) * nCols]
#define B_(i,j) B [(i) + (const size_t)(j) * nCols]
double* R1 = (double*)malloc1(nCols * nCols * sizeof(double), // nCols rows of R
"R1\t\t\tnCols %d nCols %d sizeof(double) %d",
nCols, nCols, sizeof(double));
double* B = (double*)calloc1(nCols * nCols, sizeof(double), // rhs of R1 * x = B
"B\t\t\tnCols %d nCols %d sizeof(double) %d",
nCols, nCols, sizeof(double));
int i, j;
for(i = 0; i < nCols; i++) { // copy nCols rows of R into R1
for(j = 0; j < nCols; j++)
R1_(i,j) = R_(i,j);
B_(i,i) = 1; // set diag of B to 1
}
int job = 1; // 1 means solve R1 * x = B where R1 is upper triangular
int info = 0;
for(i = 0; i < nCols; i++) {
dtrsl_( // LINPACK function
R1, // in: t, matrix of the system, untouched
(int*)&nCols, // in: ldt (typecast discards const)
(int*)&nCols, // in: n
&B_(0,i), // io: b, on return has solution x
&job, // in:
&info); // io:
ASSERT(info == 0);
}
// B is now inverse(R1). Calculate B x B.
for(i = 0; i < nCols; i++)
for(j = 0; j < nCols; j++) {
double Sum = 0;
for(int k = max(i,j); k < nCols; k++)
Sum += B_(i,k) * B_(j,k);
B_(i,j) = B_(j,i) = Sum;
}
for(i = 0; i < nCols; i++)
Diags[i] = B_(i,i);
free1(B);
free1(R1);
}
//-----------------------------------------------------------------------------
// Regress y on the used columns of x, in the standard way (using QR).
// UsedCols[i] is true for each each used col i in x; unused cols are ignored.
//
// The returned Betas argument is computed from, and is indexed on,
// the compacted x vector, not on the original x.
//
// The returned iPivots should only be used if *pnRank != nUsedCols.
// The entries of iPivots refer to columns in the full x (and are 0 based).
// Entries in iPivots at *pnRank and above specify linearly dependent columns in x.
//
// To maximize compatibility we call the same routines as the R function lm.
static void Regress(
double Betas[], // out: nUsedCols * nResp, can be NULL
double Residuals[], // out: nCases * nResp, can be NULL
double* pRss, // out: RSS, summed over all nResp, can be NULL
double Diags[], // out: diags of inv(transpose(x) * x), can be NULL
int* pnRank, // out: nbr of indep cols in x, can be NULL
int iPivots[], // out: nCols, can be NULL
const double x[], // in: nCases x nCols, must include intercept
const double y[], // in: nCases x nResp
const size_t nCases, // in: number of rows in x and in y
const int nResp, // in: number of cols in y
const int nCols, // in: number of columns in x, some may not be used
const bool UsedCols[]) // in: specifies used columns in x
{
double* xUsed;
int nUsedCols = CopyUsedCols(&xUsed, x, nCases, nCols, UsedCols);
bool MustFreeResiduals = false;
if(Residuals == NULL) {
Residuals = (double*)malloc1(nCases * nResp * sizeof(double),
"Residuals\t\tnCases %d nResp %d sizeof(double) %d",
(int)nCases, nResp, sizeof(double));
MustFreeResiduals = true;
}
bool MustFreePivots = false;
if(iPivots == NULL) {
iPivots = (int*)malloc1(nUsedCols * sizeof(int),
"iPivots\t\tnUsedCols %d sizeof(int) %d",
nUsedCols, sizeof(int));
MustFreePivots = true;
}
int iCol;
for(iCol = 0; iCol < nUsedCols; iCol++)
iPivots[iCol] = iCol+1;
double* qraux = (double*)malloc1(nUsedCols * sizeof(double),
"qraux\t\t\tnUsedCols %d sizeof(double) %d",
nUsedCols, sizeof(double));
// work size must be nUsedCols*2 for dqrdc2, and nCases*nUsedCols for
// dqrsl where it is used for qy, qty, and rsd
double* work = (double*)malloc1(
max((size_t)nUsedCols * 2, nCases * nUsedCols) * sizeof(double),
"work\t\t\tnCases %d nUsedCols %d sizeof(double) %d",
(int)nCases, nUsedCols, sizeof(double));
int nCases1 = (int)nCases; // type convert from size_t
int nRank;
dqrdc2_( // R function, QR decomp based on LINPACK dqrdc
xUsed, // io: x, on return upper tri of x is R of QR
&nCases1, // in: ldx
&nCases1, // in: n
&nUsedCols, // in: p
(double*)&QR_TOL, // in: tol
&nRank, // out: k, num of indep cols of x
qraux, // out: qraux
iPivots, // out: jpvt
work); // work
double Rss = 0;
const bool NeedResiduals = !MustFreeResiduals || pRss;
int job = (Betas? 100: 0) + (NeedResiduals? 10: 0);
if(job) { // job will be zero if all we need are the iPivots from dqrdc2
for(int iResp = 0; iResp < nResp; iResp++) {
int info;
dqrsl_( // LINPACK function
xUsed, // in: x, generated by dqrdc2
&nCases1, // in: ldx
&nCases1, // in: n
&nRank, // in: k
qraux, // in: qraux
(double*)(y + iResp * nCases), // in: y
NULL, // out: qy, unused here
work, // out: qty, required if rsd in job
Betas? // out: b, only needed if user asked for them
(double*)(&Betas_(0,iResp)): work,
NeedResiduals? // out: rsd
(double*)(&Residuals_(0,iResp)): NULL,
NULL, // out: xb = yHat, unused here
&job, // in: job
&info); // in: info
ASSERT(info == 0);
// compute Residuals and Rss (sum over all responses)
if(NeedResiduals)
for(int i = 0; i < (const int)nCases; i++)
Rss += sq(Residuals_(i, iResp));
}
if(pRss)
*pRss = Rss;
}
if(nRank != nUsedCols &&
!MustFreePivots ) { // only bother if caller wants iPivots back
// adjust iPivots for missing cols in UsedCols and for 1 offset
int* PivotOffset = (int*)malloc1(nCols * sizeof(int),
"PivotOffset\t\t\tnCols %d sizeof(int) %d",
nCols, sizeof(int));
int nOffset = 0, iOld = 0;
for(iCol = 0; iCol < nCols; iCol++) {
if(!UsedCols[iCol])
nOffset++;
else {
PivotOffset[iOld] = nOffset;
if(++iOld > nUsedCols)
break;
}
}
for(iCol = 0; iCol < nUsedCols; iCol++)
iPivots[iCol] = iPivots[iCol] - 1 + PivotOffset[iPivots[iCol] - 1];
free1(PivotOffset);
}
if(pnRank)
*pnRank = nRank;
if(Diags)
CalcDiags(Diags, xUsed, nCases, nUsedCols);
if(MustFreePivots)
free1(iPivots);
if(MustFreeResiduals)
free1(Residuals);
free1(xUsed);
free1(qraux);
free1(work);
}
//-----------------------------------------------------------------------------
// This routine is for testing Regress from R, to compare results to R's lm().
#if USING_R
void RegressR( // for testing earth routine Regress from R
double Betas[], // out: (nUsedCols+1) * nResp, +1 is for intercept
double Residuals[], // out: nCases * nResp
double Rss[], // out: RSS, summed over all nResp
double Diags[], // out: diags of inv(transpose(x) * x)
int* pnRank, // out: nbr of indep cols in x
int iPivots[], // out: nCols
const double x[], // in: nCases x nCols
const double y[], // in: nCases x nResp
const int* pnCases, // in: number of rows in x and in y
const int* pnResp, // in: number of cols in y
int* pnCols, // in: number of columns in x, some may not be used
const bool UsedCols[]) // in: specifies used columns in x
{
const size_t nCases1 = *pnCases; // type convert
Regress(Betas, Residuals, Rss, Diags, pnRank, iPivots,
x, y, nCases1, *pnResp, *pnCols, UsedCols);
}
#endif
//-----------------------------------------------------------------------------
// Regress y on bx to get Residuals and Betas. If bx isn't of full rank,
// remove dependent cols, update UsedCols, and regress again on the bx with
// removed cols.
static void RegressAndFix(
double Betas[], // out: nMaxTerms x nResp, can be NULL
double Residuals[], // out: nCases x nResp, can be NULL
double Diags[], // out: if !NULL set to diags of inv(transpose(bx) * bx)
bool UsedCols[], // io: will remove cols if necessary, nMaxTerms x 1
const double bx[], // in: nCases x nMaxTerms
const double y[], // in: nCases x nResp
const size_t nCases, // in
const int nResp, // in: number of cols in y
const int nTerms) // in: number of cols in bx, some may not be used
{
int nRank;
int* iPivots = (int*)malloc1(nTerms * sizeof(int),
"iPivots\t\tnTerms %d sizeof(int) %d",
nTerms, sizeof(int));
Regress(Betas, Residuals, NULL, Diags, &nRank, iPivots,
bx, y, nCases, nResp, nTerms, UsedCols);
int nUsedCols = GetNbrUsedCols(UsedCols, nTerms);
const int nDeficient = nUsedCols - nRank;
if(nDeficient) { // rank deficient?
// Remove linearly dependent columns.
// The lin dep columns are at index nRank and higher in iPivots.
for(int iCol = nRank; iCol < nUsedCols; iCol++)
UsedCols[iPivots[iCol]] = false;
Regress(Betas, Residuals, NULL, Diags, &nRank, NULL,
bx, y, nCases, nResp, nTerms, UsedCols);
nUsedCols = nUsedCols - nDeficient;
if(nRank != nUsedCols)
warning("Could not fix rank deficient bx: nUsedCols %d nRank %d",
nUsedCols, nRank);
else {
tprintf(1,
"Fixed rank deficient bx by removing %d term%s, %d term%s remain%s\n",
nDeficient, ((nDeficient==1)? "": "s"),
nUsedCols, ((nUsedCols==1)? "": "s"), ((nUsedCols==1)? "s": ""));
tprintf(4, "\n");
}
}
free1(iPivots);
}
//-----------------------------------------------------------------------------
static INLINE double Mean(const double x[], size_t n)
{
double mean = 0;
for(size_t i = 0; i < n; i++)
mean += x[i] / n;
return mean;
}
//-----------------------------------------------------------------------------
// get mean centered sum of squares
static INLINE double SumOfSquares(const double x[], const double mean, size_t n)
{
double ss = 0;
for(size_t i = 0; i < n; i++)
ss += sq(x[i] - mean);
return ss;
}
//-----------------------------------------------------------------------------
static INLINE double GetGcv(const int nTerms, // nbr basis terms including intercept
const size_t nCases, double Rss, const double Penalty)
{
double Cost;
if(Penalty == -1) // special case: terms and knots are free
Cost = 0;
else {
const double nKnots = ((double)nTerms-1) / 2;
Cost = (nTerms + Penalty * nKnots) / nCases;
}
// test against Cost ensures that GCVs are non-decreasing as nbr of terms increases
return Cost >= 1? POS_INF : Rss / (nCases * sq(1 - Cost));
}
//-----------------------------------------------------------------------------
// Check if model term type is already in model, to avoid a linear dependence.
// TODO The code in this routine doesn't seem to make sense.
static bool GetNewFormFlag(const int iPred, const int iTerm,
const int Dirs[], const bool UsedCols[],
const int nTerms, const int nPreds, const int nMaxTerms)
{
bool IsNewForm = true;
for(int i = 1; i < nTerms; i++) // start at 1 to skip intercept
if(UsedCols[i]) {
IsNewForm = false;
if(Dirs_(i,iPred) == 0) // unused in term i
return true;
// TODO if the following code is commented out, test suite passes (!)
for(int j = 0; j < nPreds; j++)
if(j != iPred && (Dirs_(i,j) != 0) != (Dirs_(iTerm,j) != 0))
return true;
}
return IsNewForm;
}
//-----------------------------------------------------------------------------
static double GetCut(
const int i,
const int iPred,
const size_t nCases,
const double x[],
const int xOrder[])
{
if(i < 0 || i >= (const int)nCases)
error("GetCut i %d: i < 0 || i >= nCases", i);
const int ix = xOrder_(i,iPred);
if(ix < 0 || ix >= (const int)nCases)
error("GetCut ix %d: ix < 0 || ix >= nCases", ix);
return x_(ix,iPred);
}
//-----------------------------------------------------------------------------
// The BetaCache is used when searching for a new term pair, via FindTerm.
// Most of the calculation for the orthogonal regression betas is repeated
// with the same data, and thus we can save time by caching betas.
// (The "Betas" are the regression coefficients.)
//
// iParent is the term that forms the base for the new term
// iPred is the predictor for the new term
// iOrthCol is the column index in the bxOrth matrix
static double* BetaCacheGlobal; // [iOrthCol,iParent,iPred]
// dim nPreds x nMaxTerms x nMaxTerms
static void InitBetaCache(const bool UseBetaCache,
const int nMaxTerms, const int nPreds)
{
int nCache = nMaxTerms * nMaxTerms * nPreds;
if(!UseBetaCache) {
BetaCacheGlobal = NULL;
// 3e9 below is somewhat arbitrary but seems about right (in 2011)
} else if(nCache * sizeof(double) > 3e9) {
printf(
"\nNote: earth's beta cache would require %s, so forcing Use.beta.cache=FALSE.\n"
" Invoke earth with Use.beta.cache=FALSE to make this message go away.\n\n",
sFormatMemSize(nCache * sizeof(double), false));
BetaCacheGlobal = NULL;
} else {
tprintf(5, "BetaCache %s\n", // print cache size
sFormatMemSize(nCache * sizeof(double), false));
BetaCacheGlobal = (double*)malloc1(nCache * sizeof(double),
"BetaCacheGlobal\tnMaxTerms %d nMaxTerms %d nPreds %d sizeof(double) %d",
nMaxTerms, nMaxTerms, nPreds, sizeof(double));
for(int i = 0; i < nCache; i++) // mark all entries as uninitialized
BetaCacheGlobal[i] = POS_INF;
}
}
static void FreeBetaCache(void)
{
if(BetaCacheGlobal)
free1(BetaCacheGlobal);
}
//-----------------------------------------------------------------------------
// Init a new bxOrthCol to the residuals from regressing y on the used columns
// of the orthogonal matrix bxOrth. The length (i.e. sum of squares divided
// by nCases) of each column of bxOrth must be 1 with mean 0 (except the
// first column which is the intercept).
//
// In practice this function is called with the params shown in {braces}
// and is called only by InitBxOrthCol.
//
// This function must be fast.
//
// In calculation of Beta, we used to have
// xty += pbxOrth[i] * y[i];
// and now we have
// xty += pbxOrth[i] * bxOrthCol[i];
// i.e. we use the "modified" instead of the "classic" Gram Schmidt.
// This is less susceptible to numerical error, although it is rare
// to see the effect in practice in earth models (but you can see it in the
// final model in the test suite function test.zigzag in test.weights.R).
static INLINE void OrthogResiduals(
double bxOrthCol[], // out: nCases x 1 { bxOrth[,nTerms] }
const double y[], // in: nCases x nResp { bx[,nTerms], xbx }
const double bxOrth[], // in: nTerms x nPreds { bxOrth }
const size_t nCases, // in
const int nTerms, // in: nTerms in model, i.e. number of used cols in bxOrth
const bool UsedTerms[], // in: UsedTerms[i] is true if col is used, unused cols ignored
// Following parameters are only for the beta cache
const int iParent, // in: if >= 0, use BetaCacheGlobal {FindTerm iTerm, addTermP -1}
const int iPred, // in: predictor index i.e. col index in input matrix x
const int nMaxTerms) // in:
{
double* pCache;
if(iParent >= 0 && BetaCacheGlobal)
pCache = BetaCacheGlobal + iParent*nMaxTerms + iPred*sq(nMaxTerms);
else
pCache = NULL;
memcpy(bxOrthCol, y, nCases * sizeof(double));
for(int iTerm = 0; iTerm < nTerms; iTerm++)
if(UsedTerms[iTerm]) {
const double* pbxOrth = &bxOrth_(0, iTerm);
double Beta;
if(pCache && pCache[iTerm] != POS_INF)
Beta = pCache[iTerm];
else {
double xty = 0;
for(int i = 0; i < (const int)nCases; i++)
xty += pbxOrth[i] * bxOrthCol[i]; // see header comment
Beta = xty; // no need to divide by xtx, it is 1
ASSERT(FINITE(Beta));
if(pCache)
pCache[iTerm] = Beta;
}
if(USE_BLAS) {
const double NegBeta = -Beta;
const int nCases1 = (int)nCases; // type convert from size_t
daxpy_(&nCases1, &NegBeta, pbxOrth, &ONE, bxOrthCol, &ONE);
} else for(int i = 0; i < (const int)nCases; i++)
bxOrthCol[i] -= Beta * pbxOrth[i];
}
}
//-----------------------------------------------------------------------------
// Init the rightmost column of bxOrth i.e. the column indexed by nTerms.
// The new col is the normalized residuals from regressing y on the
// lower (i.e. already existing) cols of bxOrth.
// Also updates bxOrthCenteredT and bxOrthMean.
//
// In practice this function is called only with the params shown in {braces}
static INLINE void InitBxOrthCol(
double bxOrth[], // io: col nTerms is changed, other cols not touched
double bxOrthCenteredT[], // io: kept in sync with bxOrth
double bxOrthMean[], // io: element at nTerms is updated
bool* pGoodCol, // io: true if col sum-of-squares is greater than MIN_BX_SOS
const double* y, // in: { AddCandLinTerm xbx, AddTermPair bx[,nTerms] }
const int nTerms, // in: column goes in at index nTerms, 0 is the intercept
const bool WorkingSet[], // in
const size_t nCases, // in
const int nMaxTerms, // in
const int iCacheTerm, // in: if >= 0, use BetaCacheGlobal {FindTerm iTerm, AddTermP -1}
// if < 0 then recalc Betas from scratch
const int iPred) // in: predictor index i.e. col index in input matrix x
{
*pGoodCol = true;
int i;
if(nTerms == 0) { // column 0, the intercept
double len = 1 / sqrt((double)nCases);
for(i = 0; i < (const int)nCases; i++)
bxOrth_(i,0) = len;
bxOrthMean[0] = len;
} else if(nTerms == 1) { // column 1, the first basis function, y = xbx = x[,1]
double yMean = Mean(y, nCases);
for(i = 0; i < (const int)nCases; i++)
bxOrth_(i,1) = y[i] - yMean;
} else
OrthogResiduals(&bxOrth_(0,nTerms), // resids go in rightmost col of bxOrth at nTerms
y, bxOrth, nCases, nTerms, WorkingSet, iCacheTerm, iPred, nMaxTerms);
if(nTerms > 0) {
// normalize the column to length 1 and init bxOrthMean[nTerms]
double bxOrthSS = SumOfSquares(&bxOrth_(0,nTerms), 0, nCases);
if(bxOrthSS <= MIN_BX_SOS)
*pGoodCol = false;
// iCacheTerm will be negative unless called by AddCandidateLinearTerm
if(bxOrthSS > (iCacheTerm < 0? 0: MIN_BX_SOS)) {
bxOrthMean[nTerms] = Mean(&bxOrth_(0,nTerms), nCases);
const double len = sqrt(bxOrthSS);
for(i = 0; i < (const int)nCases; i++)
bxOrth_(i,nTerms) /= len;
} else {
bxOrthMean[nTerms] = 0;
memset(&bxOrth_(0,nTerms), 0, nCases * sizeof(double));
}
}
for(i = 0; i < (const int)nCases; i++) // keep bxOrthCenteredT in sync
bxOrthCenteredT_(nTerms,i) = bxOrth_(i,nTerms) - bxOrthMean[nTerms];
}
//-----------------------------------------------------------------------------
// Add a new term pair to the arrays.
// Each term in the new term pair is a copy of an existing parent term but extended
// by multiplying it by a new hinge function at the selected knot.
// If the upper term in the term pair is invalid then we still add the upper
// term but mark it as false in FullSet.
static void AddTermPair(
int Dirs[], // io
double Cuts[], // io
double bx[], // io: MARS basis matrix
double bxOrth[], // io
double bxOrthCenteredT[], // io
double bxOrthMean[], // io
bool FullSet[], // io
bool* pIsNewForm, // io
int nDegree[], // io: degree of each term, degree of intercept is 0
int nUses[], // io: nbr of times each predictor is used in the model
const int nTerms, // in: new term pair goes in at index nTerms and nTerms1
const int iBestParent, // in: parent term
const int iBestCase, // in
const int iBestPred, // in
const int nPreds, // in
const size_t nCases, // in
const int nMaxTerms, // in
const bool LinPredIsBest, // in: true if pred should enter linearly (no knot)
const int LinPreds[], // in: user specified preds which must enter linearly
const double x[], // in
const int xOrder[], // in
const bool Weighted) // in
{
const int nTerms1 = nTerms+1;
// copy the parent term to the new term pair
int iPred;
for(iPred = 0; iPred < nPreds; iPred++) {
Dirs_(nTerms, iPred) = Dirs_(nTerms1,iPred) = Dirs_(iBestParent,iPred);
Cuts_(nTerms, iPred) = Cuts_(nTerms1,iPred) = Cuts_(iBestParent,iPred);
}
// incorporate the new hinge function
nDegree[nTerms] = nDegree[nTerms1] = nDegree[iBestParent] + 1;
int DirEntry = 1;
if(LinPreds[iBestPred] || LinPredIsBest) { // changed in earth 4.0.0
ASSERT(LinPredIsBest);
DirEntry = 2;
}
Dirs_(nTerms, iBestPred) = DirEntry;
Dirs_(nTerms1,iBestPred) = -1; // will be ignored if adding only one hinge
const double BestCut = GetCut(iBestCase, iBestPred, nCases, x, xOrder);
Cuts_(nTerms, iBestPred) = Cuts_(nTerms1,iBestPred) = BestCut;
// Fill in new columns of bx, at nTerms and nTerms+1 (left and right hinges).
#if WEIGHTS
// in FindWeightedPredGivenParent, we used the two columns
// in bx as a scratch buffer, so zero them again
ASSERT(nTerms+1 < nMaxTerms);
memset(&bx_(0,nTerms), 0, nCases * sizeof(double));
memset(&bx_(0,nTerms+1), 0, nCases * sizeof(double));
#endif
int i;
if(DirEntry == 2) { // linpred?
for(i = 0; i < (const int)nCases; i++) // add single term
bx_(i,nTerms) = bx_(i,iBestParent) * x_(i,iBestPred);
} else for(i = 0; i < (const int)nCases; i++) { // add term pair
const int iOrdered = xOrder_(i, iBestPred);
const double xi = x_(iOrdered, iBestPred);
if(i > iBestCase)
bx_(iOrdered, nTerms) = bx_(iOrdered, iBestParent) * (xi - BestCut);
else
bx_(iOrdered, nTerms1) = bx_(iOrdered, iBestParent) * (BestCut - xi);
}
nUses[iBestPred]++;
// init the col in bxOrth at nTerms and init bxOrthMean[nTerms]
FullSet[nTerms] = true;
bool GoodCol;
InitBxOrthCol(bxOrth, bxOrthCenteredT, bxOrthMean, &GoodCol,
&bx_(0,nTerms), nTerms, FullSet, nCases, nMaxTerms, -1, nPreds);
// -1 means don't use BetaCacheGlobal, calc Betas afresh
#if 0 // TODO commented out because this sometimes happens
if(!GoodCol)
printf("GoodCol is false in AddTermPair\n");
#endif
// init the col in bxOrth at nTerms1 and init bxOrthMean[nTerms1]
if(!LinPredIsBest && *pIsNewForm)
FullSet[nTerms1] = true;
if(FullSet[nTerms1]) {
InitBxOrthCol(bxOrth, bxOrthCenteredT, bxOrthMean, &GoodCol,
&bx_(0,nTerms1), nTerms1, FullSet, nCases, nMaxTerms, -1, iPred);
if(Weighted && !GoodCol) {
// !GoodCol usuall happens only when weights are used. For the
// nonweighted situation we would have already cleared IsNewForm in
// FindPredGivenParent (inside AddCandidateLinearTerm), although
// we sometimes get a non GoodCol here without weights.
tprintf(6, "clear IsNewForm\n");
*pIsNewForm = false;
FullSet[nTerms1] = false;
}
}
if(!FullSet[nTerms1]) {
// If the term is not valid, then we don't wan't to use it as the
// base for a new term later (in FindTerm). Enforce this by setting
// nDegree to a value greater than any posssible nMaxDegree.
nDegree[nTerms1] = MAX_DEGREE + 1;
memset(&bxOrth_(0,nTerms1), 0, nCases * sizeof(double));
bxOrthMean[nTerms1] = 0;
for(i = 0; i < (const int)nCases; i++) // keep bxOrthCenteredT in sync
bxOrthCenteredT_(nTerms1,i) = 0;
}
}
//-----------------------------------------------------------------------------
static int GetNbrUsed( // Nm in Friedman's notation
const size_t nCases, // in
const int iParent, // in
const double bx[]) // in: MARS basis matrix
{
int nUsed = 0;
if(bx == NULL)
nUsed = (const int)nCases;
else for(int i = 0; i < (const int)nCases; i++)
if(bx_(i,iParent) > 0)
nUsed++;
return nUsed;
}
//-----------------------------------------------------------------------------
static int GetEndSpan(
const int nPreds,
const int nDegree, // in: degree of current term (for adjusting endspan)
const size_t nCases)
{
int nEndSpan = 1;
if(nEndSpanGlobal > 0) // user specified endspan?
nEndSpan = nEndSpanGlobal;
else if(nEndSpanGlobal == 0) { // auto?
// eqn 45 FriedmanMars (see refs)
static const double log_2 = 0.69315; // log(2)
static const double temp1 = 7.32193; // 3 + log(20)/log(2);
nEndSpan = (int)(temp1 + log((double)nPreds) / log_2);
} else // negative endspan illegal
error("endspan %d < 0", nEndSpanGlobal);
if(nDegree >= 2) // .5 below makes (int) cast act as round
nEndSpan += (int)(AdjustEndSpanGlobal * nEndSpan + .5);
if(nEndSpan > (const int)nCases / 2 - 1) // always at least one knot, so above adjustment
nEndSpan = (const int)nCases / 2 - 1; // doesn't completely inhibit degree2 terms
nEndSpan = max(1, nEndSpan);
return nEndSpan;
}
//-----------------------------------------------------------------------------
static void GetSpanParams(
int* pnMinSpan, // out: number cases between knots
int* pnEndSpan, // out: number of cases ignored on each end
int* pnStartSpan, // out: number of cases from end until first knot
const size_t nCases, // in
const int nPreds, // in
const int nDegree, // in: degree of current term (for adjusting endspan)
const int iParent, // in
const double bx[]) // in: MARS basis matrix, can be NULL
{
const int nEndSpan = GetEndSpan(nPreds, nDegree, nCases);
int nStartSpan = 0, nMinSpan = 0;
if(nMinSpanGlobal < 0) { // treat negative minspan as number of knots
// get nMinSpan
nMinSpan = (int)(ceil(nCases / (1.-nMinSpanGlobal))); // convert nknots to minspan
// get nStartSpan
nStartSpan = nMinSpan;
while(nStartSpan < nEndSpan)
nStartSpan += nMinSpan;
nStartSpan--;
nStartSpan = max(1, nStartSpan);
} else {
// get nMinSpan
if(nMinSpanGlobal > 0) // user specified minspan?
nMinSpan = nMinSpanGlobal;
else if(nMinSpanGlobal == 0) { // auto?
// eqn 43 in FriedmanMars paper (see refs)
const int nUsed = GetNbrUsed(nCases, iParent, bx); // Nm in Friedmans notation
static const double temp1 = 2.9702; // -log(-log(0.95)
static const double temp2 = 1.7329; // 2.5 * log(2)
nMinSpan = (int)((temp1 + log((double)(nPreds * nUsed))) / temp2);
}
nMinSpan = max(1, nMinSpan);
// get nStartSpan
const int nAvail = max(0, (const int)nCases - 2 * nEndSpan);
nStartSpan = nAvail / 2; // if space for only one knot, put it in center
if(nAvail > nMinSpan) { // space for more than one knot?
const int nDiv = nAvail / nMinSpan;
if(nAvail == nDiv * nMinSpan)
nStartSpan = nMinSpan / 2;
else
nStartSpan = (nAvail - nDiv * nMinSpan) / 2;
}
// TODO consider moving the following line of code into the "}" above
nStartSpan = max(1, nEndSpan + nStartSpan);
}
*pnStartSpan = nStartSpan;
*pnMinSpan = nMinSpan;
*pnEndSpan = nEndSpan;
}
//-----------------------------------------------------------------------------
// The caller has selected a candidate predictor iPred and a candidate iParent.
// This function now selects a knot. If it finds a knot it will
// update *piBestCase and pRssDeltaForParPredPair.
//
// The general idea: scan backwards through all (ordered) values (i.e. potential
// knots) for the given predictor iPred, calculating RssDelta.
// If RssDelta > *pRssDeltaForParPredPair (and all else is ok), then
// select the knot (by updating *piBestCase and *pRssDeltaForParPredPair).
//
// We want to add a term pair at index iNewCol and iNewCol+1.
// There are currently nTerms in the model.
//
// This function must be fast.
static INLINE void FindKnot(
int* piBestCase, // out: possibly updated, row index in x
double* pRssDeltaForParPredPair, // io: updated if knot is better
double CovCol[], // scratch buffer, overwritten, nTerms x 1
double ycboSum[], // scratch buffer, overwritten, nMaxTerms x nResp
double CovSx[], // scratch buffer, overwritten, nTerms x 1
double* ybxSum, // scratch buffer, overwritten, nResp x 1
const int iNewCol, // in: tentative knot goes into bx[iNewCol]
const int iParent, // in: parent term
const int iPred, // in: predictor index
const size_t nCases, // in
const int nResp, // in: number of cols in y
const int nMaxTerms, // in
const double RssDeltaLin, // in: change in RSS if predictor iPred enters linearly
const double MaxLegalRssDelta, // in: FindKnot rejects any changes in Rss greater than this
const double bx[], // in: MARS basis matrix
const double bxOrth[], // in
const double bxOrthCenteredT[], // in
const double bxOrthMean[], // in
const double x[], // in: nCases x nPreds
const double y[], // in: nCases x nResp
const int xOrder[], // in
const double yMean[], // in: vector nResp x 1
const int nStartSpan, // in: number of cases from end until first knot
const int nMinSpan, // in: number cases between knots
const int nEndSpan, // in: number of cases ignored on each end
const double NewVarAdjust, // in: 1 if not a new var, 1/(1+NewVarPenalty) if new var
const double RssBeforeNewTerm) // in: used only when trace >= 8
{
tprintf(8, "--FindKnotBegin-- iPred %d iNewCol %d RssBeforeAddingHinge %g "
"nMinSpan %d nEndSpan %d nStartSpan %d\n",
iPred+IOFFSET, iNewCol+IOFFSET, RssBeforeNewTerm,
nMinSpan, nEndSpan, nStartSpan);
ASSERT(MaxLegalRssDelta > 0);
// Tol was .01 prior to earth 4.4.0 but that caused zigzag runout in test.weights.R.
// The comparison against iNewCol provides some measure of back compatibility and
// helps prevent overfitting in small models. The 15 is fairly arb but was chosen
// to keep small models from having nearby knots, for the zizgag function in
// test.weights.R, and for the spkmap neural data.
// TODO This isn't a clean solution.
const double Tol = iNewCol < 15? .01 : 1e-5;
int iResp;
for(iResp = 0; iResp < nResp; iResp++)
ycboSum_(iNewCol, iResp) = 0;
memset(CovCol, 0, (iNewCol+1) * sizeof(double));
memset(CovSx, 0, (iNewCol+1) * sizeof(double));
memset(ybxSum, 0, nResp * sizeof(double));
double bxSum = 0, bxSqSum = 0, bxSqxSum = 0, bxxSum = 0, st = 0;
int iSpan = nStartSpan;
for(int i = (const int)nCases-2; i >= nEndSpan; i--) { // -2 allows for ix1
// may Mars have mercy on the poor soul who enters here
const int ix0 = xOrder_(i, iPred); // get the x's in descending order
const double x0 = x_(ix0,iPred); // the knot (printed as Cut in trace prints)
const int ix1 = xOrder_(i+1, iPred);
const double x1 = x_(ix1, iPred); // case next to the cut
const double bx1 = bx_(ix1, iParent);
const double bxSq = sq(bx1);
const double xDelta = x1 - x0; // will always be non negative
if(USE_BLAS) {
daxpy_(&iNewCol, &bx1, &bxOrthCenteredT_(0,ix1), &ONE, CovSx, &ONE);
daxpy_(&iNewCol, &xDelta, CovSx, &ONE, CovCol, &ONE);
} else for(int it = 0; it < iNewCol; it++) {
CovSx[it] += (bxOrth_(ix1,it) - bxOrthMean[it]) * bx1;
CovCol[it] += xDelta * CovSx[it];
}
bxSum += bx1;
bxSqSum += bxSq;
bxxSum += bx1 * x1;
bxSqxSum += bxSq * x1;
const double su = st;
st = bxxSum - bxSum * x0;
CovCol[iNewCol] += xDelta * (2 * bxSqxSum - bxSqSum * (x0 + x1)) +
(sq(su) - sq(st)) / nCases;
if(nResp == 1) { // treat nResp==1 as a special case, for speed
ybxSum[0] += (y_(ix1, 0) - yMean[0]) * bx1;
ycboSum_(iNewCol, 0) += xDelta * ybxSum[0];
} else for(iResp = 0; iResp < nResp; iResp++) {
ybxSum[iResp] += (y_(ix1, iResp) - yMean[iResp]) * bx1;
ycboSum_(iNewCol, iResp) += xDelta * ybxSum[iResp];
}
if(bx1 > 0 && CovCol[iNewCol] > 0 && --iSpan == 0) {
iSpan = nMinSpan;
double RssDelta = 0;
bool Best = false, TolGood = true;
// calculate RssDelta and see if this knot beats the previous best
RssDelta = 0;
double temp1, temp2;
for(iResp = 0; iResp < nResp; iResp++) {
if(USE_BLAS) {
temp1 =
ycboSum_(iNewCol,iResp) -
ddot_(&iNewCol, &ycboSum_(0,iResp), &ONE, CovCol, &ONE);
temp2 =
CovCol[iNewCol] - ddot_(&iNewCol, CovCol, &ONE, CovCol, &ONE);
} else {
temp1 = ycboSum_(iNewCol,iResp);
temp2 = CovCol[iNewCol];
for(int it = 0; it < iNewCol; it++) {
temp1 -= ycboSum_(it,iResp) * CovCol[it];
temp2 -= sq(CovCol[it]);
}
}
// TODO HastieTibs code has a comment saying the following has to be fixed
if(temp2 / CovCol[iNewCol] > Tol)
RssDelta += sq(temp1) / temp2;
else
TolGood = false;
}
RssDelta = NewVarAdjust * (RssDeltaLin + RssDelta);
if(RssDelta > *pRssDeltaForParPredPair &&
RssDelta < MaxLegalRssDelta) {
*piBestCase = i;
*pRssDeltaForParPredPair = RssDelta;
Best = true;
}
if(TraceGlobal >= 8) {
const double RssWithKnot = RssBeforeNewTerm - RssDelta;
printf("--FindKnot--Case %4d RssWithKnot %12.5g RssDelta %12.5g "
"Cut % 12.5g ",
i+IOFFSET, MaybeZero(RssWithKnot), MaybeZero(RssDelta), x0);
// separate trace to make it easier to check compat
// with FindWeightedKnot when trace==8
tprintf(9, "bx1G %d CovColG %d TolG %d MaxG %d ",
bx1 > 0, CovCol[iNewCol] > 0, TolGood,
RssDelta < MaxLegalRssDelta);
printf("%s\n", Best? " best": "");
}
} else if(TraceGlobal >= 9)
printf("--FindKnot--Case %4d iSpan %d bx1 % 8.4f\n",
i+IOFFSET, iSpan, bx1);
} // for
tprintf(8, "--FindKnotEnd--\n");
}
//-----------------------------------------------------------------------------
// Add a candidate term at bx[,nTerms], with the parent term multiplied by
// the predictor iPred entering linearly. Do this by setting the knot at
// the lowest value xMin of x, since max(0,x-xMin)==x-xMin for all x. The
// change in RSS caused by adding this term forms the base RSS delta which
// we will try to beat in the search in FindKnot.
//
// This also initializes CovCol, bxOrth[,nTerms], and ycboSum[nTerms,]
static INLINE void AddCandidateLinearTerm(
double* pRssDeltaLin, // out: change to RSS caused by adding new term
bool* pIsNewForm, // out: true on entry, may be cleared by InitBxOrthCol
double xbx[], // out: nCases x 1
double CovCol[], // out: nMaxTerms x 1
double ycboSum[], // io: nMaxTerms x nResp
double bxOrth[], // io
double bxOrthCenteredT[], // io
double bxOrthMean[], // io
const int iPred, // in
const int iParent, // in
const double x[], // in: nCases x nPreds
const double y[], // in: nCases x nResp, scaled y
const size_t nCases, // in
const int nResp, // in: number of cols in y
const int nTerms, // in
const int nMaxTerms, // in
const double bx[], // in: MARS basis matrix
const bool FullSet[]) // in
{
// set xbx to x[,iPred] * bx[,iParent]
// note: when iParent==1, bx_[,iParent] is all ones, therefore xbx is x
int i;
for(i = 0; i < (const int)nCases; i++)
xbx[i] = x_(i,iPred) * bx_(i,iParent);
// Init bxOrth[,nTerms] and bxOrthMean[nTerms] for the candidate term.
// Clears both those columns, bxOrthMean, and *pIsNewForm if
// column sum-of-squares is less than MIN_BX_SOS.
InitBxOrthCol(bxOrth, bxOrthCenteredT, bxOrthMean, pIsNewForm,
xbx, nTerms, FullSet, nCases, nMaxTerms, iParent, iPred);
// init CovCol and ycboSum[nTerms], for use by FindKnot later
memset(CovCol, 0, (nTerms-1) * sizeof(double));
CovCol[nTerms] = 1;
int iResp;
for(iResp = 0; iResp < nResp; iResp++) {
ycboSum_(nTerms, iResp) = 0;
for(i = 0; i < (const int)nCases; i++)
ycboSum_(nTerms, iResp) += (y_(i, iResp) - yMean[iResp]) *
bxOrth_(i,nTerms);
}
// calculate change to RSS caused by adding candidate new term
*pRssDeltaLin = 0;
for(iResp = 0; iResp < nResp; iResp++) {
double yboSum = 0;
for(i = 0; i < (const int)nCases; i++)
yboSum += y_(i,iResp) * bxOrth_(i,nTerms);
*pRssDeltaLin += sq(yboSum);
}
}
//-----------------------------------------------------------------------------
// The caller has selected a candidate parent term iParent.
// This function now selects a predictor, and a knot for that predictor.
//
// TODO These functions have a ridiculous number of parameters, I know.
//
// TODO A note on the comparison against ALMOST_ZERO below:
// It's not a clean solution but seems to work ok.
// It was added after we saw different results on different
// machines for certain datasets e.g. (tested on earth 1.4.0)
// ldose <- rep(0:5, 2) - 2
// ldose1 <- c(0.1, 1.2, 2.3, 3.4, 4.5, 5.6, 0.3, 1.4, 2.5, 3.6, 4.7, 5.8)
// sex3 <- factor(rep(c("male", "female", "andro"), times=c(6,4,2)))
// fac3 <- factor(c("lev2", "lev2", "lev1", "lev1", "lev3", "lev3",
// "lev2", "lev2", "lev1", "lev1", "lev3", "lev3"))
// numdead <- c(1,4,9,13,18,20,0,2,6,10,12,16)
// numdead2 <- c(2,3,10,13,19,20,0,3,7,11,13,17)
// pair <- cbind(numdead, numdead2)
// df <- data.frame(sex3, ldose, ldose1, fac3)
// am <- earth(df, pair, trace=6, pmethod="none", degree=2)
static INLINE void FindPredGivenParent(
int* piBestCase, // out: untouched unless an improving term is found
int* piBestPred, // out: ditto
int* piBestParent, // out: existing term on which we are basing the new term
double* pBestRssDeltaForTerm, // io: untouched unless an improving term is found
double* pBestRssDeltaForParent, // io: used only by FAST_MARS
bool* pIsNewForm, // out
bool* pLinPredIsBest, // out: true if pred should enter linearly (no knot)
double bxOrth[], // io
double bxOrthCenteredT[], // io
double bxOrthMean[], // io
double xbx[], // io: nCases x 1
double CovSx[], // io: nMaxTerms x 1
double CovCol[], // io: nMaxTerms x 1
double ycboSum[], // io: nMaxTerms x nResp
const double bx[], // in: MARS basis matrix
const double yMean[], // in: vector nResp x 1
const double RssBeforeNewTerm, // in
const double MaxLegalRssDelta, // in: FindKnot rejects any changes in Rss greater than this
const int iParent, // in
const double x[], // in: nCases x nPreds, unweighted x
const double y[], // in: nCases x nResp, unweighted but scaled y
const size_t nCases, // in
const int nResp, // in: number of cols in y
const int nPreds, // in
const int nTerms, // in
const int nMaxTerms, // in
const bool FullSet[], // in
const int xOrder[], // in: order of each column of x array
const int nUses[], // in: nbr of times each pred is used in the model
const int Dirs[], // in
const double NewVarPenalty, // in: penalty for adding a new variable (default is 0)
const int LinPreds[], // in: nPreds x 1, 1 if predictor must enter linearly
const int nMinSpan, // in
const int nEndSpan, // in
const int nStartSpan) // in
{
double* ybxSum = (double*)malloc1(nResp * sizeof(double), // working var for FindKnot
"ybxSum\t\tnResp %d sizeof(double) %d",
nResp, sizeof(double));
bool UpdatedBestRssDelta = false;
for(int iPred = 0; iPred < nPreds; iPred++) {
tprintf(9, "\n");
if(Dirs_(iParent,iPred) != 0) { // predictor is in parent term?
tprintf(7,
"|Parent %-2d Pred %-2d %44.44s skip (pred is in parent)\n",
iParent+IOFFSET, iPred+IOFFSET, " ");
#if USING_R
} else if(!IsAllowed(iPred, iParent, Dirs, nPreds, nMaxTerms)) {
tprintf(7,
"|Parent %-2d Pred %-2d %44.44s skip (not allowed by \"allowed\" func)\n",
iParent+IOFFSET, iPred+IOFFSET, " ");
#endif
} else {
// we apply the penalty if the variable is entering for the first time
const double NewVarAdjust = 1 / (1 + (nUses[iPred] == 0? NewVarPenalty: 0));
double RssDeltaLin = 0; // change in RSS for iPred entering linearly
double UnadjustedRssDeltaLin = 0;
double RssBeforeKnot = RssBeforeNewTerm;
bool IsNewForm = GetNewFormFlag(iPred, iParent, Dirs,
FullSet, nTerms, nPreds, nMaxTerms);
if(IsNewForm) {
// Create a candidate term at bx[,nTerms],
// with iParent and iPred entering linearly
// This may clear IsNewForm.
AddCandidateLinearTerm(&UnadjustedRssDeltaLin, &IsNewForm,
xbx, CovCol, ycboSum, bxOrth, bxOrthCenteredT, bxOrthMean,
iPred, iParent, x, y,
nCases, nResp, nTerms, nMaxTerms, bx, FullSet);
RssDeltaLin = NewVarAdjust * UnadjustedRssDeltaLin;
tprintf(8,
"\n|Parent %-2d Pred %-2d Case -1 Cut % 12.4g< Rss %-12.5g",
iParent+IOFFSET, iPred+IOFFSET,
GetCut(0, iPred, nCases, x, xOrder),
MaybeZero(RssBeforeNewTerm - RssDeltaLin));
tprintf(9, " RssDeltaLin %-12.5g ", RssDeltaLin);
if(fabs(RssDeltaLin - *pBestRssDeltaForTerm) < ALMOST_ZERO) {
RssDeltaLin = *pBestRssDeltaForTerm; // see header note
tprintf(7, "RssDelta %g is ALMOST_ZERO\n",
RssDeltaLin - *pBestRssDeltaForTerm);
}
if(RssDeltaLin > *pBestRssDeltaForParent)
*pBestRssDeltaForParent = RssDeltaLin;
RssBeforeKnot -= RssDeltaLin;
if(RssDeltaLin > *pBestRssDeltaForTerm) {
// The new term (with predictor entering linearly) beats other
// candidate terms so far.
tprintf(9, "best for term (lin pred) ");
UpdatedBestRssDelta = true;
*pBestRssDeltaForTerm = RssDeltaLin;
*pLinPredIsBest = true;
*piBestCase = 0; // knot is at the lowest value of x
*piBestPred = iPred;
*piBestParent = iParent;
}
tprintf(8, "\n");
} else
tprintf(8, "|Parent %-2d Pred %-2d no new form\n",
iParent+IOFFSET, iPred+IOFFSET);
double RssDeltaForParPredPair = RssDeltaLin;
if(!LinPreds[iPred]) {
int iBestCase = -1;
FindKnot(&iBestCase, &RssDeltaForParPredPair,
CovCol, ycboSum, CovSx, ybxSum,
(IsNewForm? nTerms + 1: nTerms),
iParent, iPred, nCases, nResp, nMaxTerms,
UnadjustedRssDeltaLin, MaxLegalRssDelta,
bx, bxOrth, bxOrthCenteredT, bxOrthMean,
x, y, xOrder, yMean,
nStartSpan, nMinSpan, nEndSpan, NewVarAdjust,
RssBeforeNewTerm); // RssBeforeNewTerm is for tracing
if(RssDeltaForParPredPair > *pBestRssDeltaForParent)
*pBestRssDeltaForParent = RssDeltaForParPredPair;
if(RssDeltaForParPredPair > *pBestRssDeltaForTerm) {
UpdatedBestRssDelta = true;
*pBestRssDeltaForTerm = RssDeltaForParPredPair;
*pLinPredIsBest = false;
*piBestCase = iBestCase;
*piBestPred = iPred;
*piBestParent = iParent;
*pIsNewForm = IsNewForm;
tprintf(7,
"|Parent %-2d Pred %-2d Case %4d Cut % 12.4g Rss %-12.5g RssDelta %-12.5g%s\n",
iParent+IOFFSET, iPred+IOFFSET,
iBestCase+IOFFSET,
GetCut(iBestCase, iPred, nCases, x, xOrder),
MaybeZero(RssBeforeNewTerm - *pBestRssDeltaForTerm),
MaybeZero(*pBestRssDeltaForTerm),
RssDeltaForParPredPair > *pBestRssDeltaForTerm?
" best for term": "");
} else
tprintf(7,
"|Parent %-2d Pred %-2d Case %4d Cut % 12.4g< Rss %-12.5g RssDelta %-12.5g\n",
iParent+IOFFSET, iPred+IOFFSET, -1,
GetCut(0, iPred, nCases, x, xOrder),
MaybeZero(RssBeforeNewTerm - RssDeltaLin),
0.);
}
} // else
} // for iPred
free1(ybxSum);
if(UpdatedBestRssDelta && NewVarPenalty != 0. && nUses[*piBestPred] == 0) {
// we applied NewVarPenalty earlier, now un-apply it
*pBestRssDeltaForTerm *= 1 + NewVarPenalty;
}
}
//-----------------------------------------------------------------------------
#if WEIGHTS
// TODO This could be made faster by caching previous results?
static INLINE void InitHinge( // return TRUE if column is not all zeros
double bxCol[], // out: this column will be initialized
const int iHinge, // in: hinge index (row in x)
const double bxParentCol[], // in: column in bx for iParent
const double xCol[], // in: column of x for iPred
const int xOrderCol[], // in: column in xOrder for iPred
const size_t nCases) // in
{
const double Cut = xCol[xOrderCol[iHinge]];
for(int i = (const int)nCases-1; i > iHinge; i--) {
const int ix = xOrderCol[i];
bxCol[ix] = bxParentCol[ix] * (xCol[ix] - Cut);
}
}
#endif
//-----------------------------------------------------------------------------
#if WEIGHTS
static double GetRegressionRss(
double x[], // io: nCases x nCols, gets overwritten
const double y[], // in: nCases x nResp
const size_t nCases, // in: number of rows in x and in y
const int nResp, // in: number of cols in y
int nCols, // in: number of columns in x
double Residuals[], // in: nCases, working storage
int iPivots[], // in: nCols, working storage
double qraux[], // in: nCols, working storage
double work[]) // in: max(nCols * 2, nCases), working storage
{
for(int iCol = 0; iCol < nCols; iCol++)
iPivots[iCol] = iCol+1;
int nCases1 = (const int)nCases; // type convert from size_t
int nRank;
dqrdc2_( // R function, QR decomp based on LINPACK dqrdc
x, // io: x, on return upper tri of x is R of QR
&nCases1, // in: ldx
&nCases1, // in: n
&nCols, // in: p
(double*)&QR_TOL, // in: tol
&nRank, // out: k, num of indep cols of x
qraux, // out: qraux
iPivots, // out: jpvt
work); // work
double Rss = 0;
int job = 10; // specify 10 because all we need are the residuals
int info;
for(int iResp = 0; iResp < nResp; iResp++) {
dqrsl_( // LINPACK function
x, // in: x, generated by dqrdc2
&nCases1, // in: ldx
&nCases1, // in: n
&nRank, // in: k
qraux, // in: qraux
(double*)(y + iResp * nCases), // in: y
NULL, // out: qy, unreferenced here
work, // out: qty, unused here but needed for dqrsl
NULL, // out: b, unreferenced here
(double*)Residuals, // out: rsd
NULL, // out: xb, unreferenced here
&job, // in: job
&info); // in: info
ASSERT(info == 0);
for(int i = 0; i < (const int)nCases; i++)
Rss += sq(Residuals[i]);
}
return Rss;
}
#endif // WEIGHTS
//-----------------------------------------------------------------------------
#if WEIGHTS
static INLINE void FindWeightedKnot(
int* piBestCase, // out: updated, row index in x
double* pRssBestKnot, // io: updated, on entry is LinRss if did lin pred
const bool UsedCols[], // in
int iNewCol, // in: tentative knot goes into bx[iNewCol]
const int iParent, // in: parent term
const int iPred, // in: predictor index
const size_t nCases, // in
const int nResp, // in: number of cols in yw
double bx[], // in: MARS basis matrix, columns nTerm and nTerm+1 filled in
const double x[], // in: nCases x nPreds, unweighted x matrix
const double yw[], // in: nCases x nResp, weighted and scaled y matrix
const int xOrder[], // in: order of each column of _unweighted_ x array
const int nStartSpan, // in: number of cases from end until first knot
const int nMinSpan, // in: number cases between knots
const int nEndSpan, // in: number of cases ignored on each end
const double RssBeforeNewTerm) // in: used only for tracing
{
tprintf(8, "--FindKnotBegin-- iPred %d iNewCol %d RssBeforeAddingHinge %g "
"nMinSpan %d nEndSpan %d nStartSpan %d\n",
iPred+IOFFSET, iNewCol+IOFFSET, RssBeforeNewTerm,
nMinSpan, nEndSpan, nStartSpan);
// we malloc everything for GetRegressionRss once here instead of in the loop
double* bxUsed;
const int nUsedCols = CopyUsedCols(&bxUsed, bx, nCases, iNewCol+1, UsedCols);
double* bxTemp = (double*)malloc1(nCases * nUsedCols * sizeof(double),
"bxTemp\t\tnCases %d nUsedCols %d sizeof(double) %d",
(const int)nCases, nUsedCols, sizeof(double));
double* Residuals = (double*)malloc1(nCases * sizeof(double),
"Residuals\t\tnCases %d sizeof(double) %d",
(const int)nCases, sizeof(double));
int* iPivots = (int*)malloc1(nUsedCols * sizeof(int),
"iPivots\t\tnUsedCols %d sizeof(int) %d",
nUsedCols, sizeof(int));
double* qraux = (double*)malloc1(nUsedCols * sizeof(double),
"qraux\t\t\tnUsedCols %d sizeof(double) %d",
nUsedCols, sizeof(double));
// in GetRegressionRss, work must be p*2 for dqrdc2, and
// nCases in dqrsl where it is used temp storage for qty
double* work = (double*)malloc1(
max(nUsedCols * 2, (const int)nCases) * sizeof(double),
"work\t\t\tnCases %d sizeof(double) %d",
(const int)nCases, sizeof(double));
// zero the current column of bxUsed, we will fill it in with the hinge functions
memset(&bxUsed_(0,iNewCol), 0, nCases * sizeof(double));
// for-loop indices and trace prints are compatible with FindKnot (when trace<=8)
int iSpan = nStartSpan;
for(int i = (const int)nCases-2; i >= nEndSpan; i--) {
// TODO This should be bx0 (not bx1)? But for compat with FindKnot we use bx1.
// In test.mods.R, using bx1 or bx0 here give almost identical results.
const int ix1 = xOrder_(i+1, iPred);
const double bx1 = bx_(ix1,iParent);
if(bx1 > 0 && --iSpan == 0) {
iSpan = nMinSpan;
bool Best = false;
InitHinge(&bxUsed_(0, iNewCol), // init this column of bxUsed
i, &bx_(0, iParent), &x_(0, iPred), &xOrder_(0, iPred), nCases);
// bxTemp is needed because GetRegressionRss destroys its first arg
memcpy(bxTemp, bxUsed, nCases * nUsedCols * sizeof(double));
const double KnotRss =
GetRegressionRss(bxTemp, yw, nCases, nResp, nUsedCols,
Residuals, iPivots, qraux, work);
// using ALMOST_ZERO here gives results closer to FindKnot
if(KnotRss < *pRssBestKnot - ALMOST_ZERO) {
*piBestCase = i;
*pRssBestKnot = KnotRss;
Best = true;
}
if(TraceGlobal >= 8) {
const double RssDelta = RssBeforeNewTerm - KnotRss;
printf(
"--FindKnot--Case %4d RssWithKnot %12.5g RssDelta %12.5g Cut % 12.5g%s\n",
i+IOFFSET, MaybeZero(KnotRss), MaybeZero(RssDelta),
GetCut(i, iPred, nCases, x, xOrder), Best? " best": "");
}
} else if(TraceGlobal >= 9)
printf("--FindKnot--Case %4d iSpan %d bx1 % 8.4f\n", i+IOFFSET, iSpan, bx1);
}
free1(work);
free1(qraux);
free1(iPivots);
free1(Residuals);
free1(bxTemp);
free1(bxUsed);
tprintf(8, "--FindKnotEnd--\n");
}
#endif // WEIGHTS
//-----------------------------------------------------------------------------
#if WEIGHTS
static INLINE void FindWeightedPredGivenParent(
int* piBestCase, // out: untouched unless an improving term is found
int* piBestPred, // out: ditto
int* piBestParent, // out: existing term on which we are basing the new term
double* pBestRssDeltaForTerm, // io: untouched unless an improving term is found
double* pBestRssDeltaForParent, // io: used only by FAST_MARS
bool* pIsNewForm, // out
bool* pLinPredIsBest, // out: true if pred should enter linearly (no knot)
double bx[], // in: MARS basis matrix
const double RssBeforeNewTerm, // in
const int iParent, // in
const double x[], // in: nCases x nPreds, unweighted x
const double yw[], // in: nCases x nResp, weighted and scaled y
const size_t nCases, // in
const int nResp, // in: number of cols in y
const int nPreds, // in
const int nTerms, // in
const int nMaxTerms, // in
const bool FullSet[], // in
const int xOrder[], // in: order of each column of _unweighted_ x array
const int nUses[], // in: nbr of times each pred is used in the model
const int Dirs[], // in
const double NewVarPenalty, // in: penalty for adding a new variable (default is 0)
const int LinPreds[], // in: nPreds x 1, 1 if predictor must enter linearly
const int nMinSpan, // in
const int nEndSpan, // in
const int nStartSpan) // in
{
if(NewVarPenalty != 0) // NewVarPenalty is not yet fully tested when weights are used
error("newvar.penalty is not yet implemented with weights");
bool UpdatedBestRssDelta = false;
bool* UsedCols = (bool*)calloc1(nMaxTerms, sizeof(bool),
"UsedCols\t\tnMaxTerms %d sizeof(bool) %d",
nMaxTerms, sizeof(bool));
for(int iTerm = 0; iTerm < nTerms; iTerm++)
UsedCols[iTerm] = true;
for(int iPred = 0; iPred < nPreds; iPred++) {
tprintf(8, "\n");
if(Dirs_(iParent,iPred) != 0) { // predictor is in parent term?
tprintf(7, "|Parent %-2d Pred %-2d %44.44s skip (pred is in parent)\n",
iParent+IOFFSET, iPred+IOFFSET, " ");
#if USING_R
} else if(!IsAllowed(iPred, iParent, Dirs, nPreds, nMaxTerms)) {
tprintf(7,
"|Parent %-2d Pred %-2d %44.44s skip (not allowed by \"allowed\" func)\n",
iParent+IOFFSET, iPred+IOFFSET, " ");
#endif
} else {
#if USING_R
// TODO we don't release UsedCols here if user interrupts
ServiceR();
#endif
// const double NewVarAdjust = 1 + (nUses[iPred] == 0? NewVarPenalty: 0);
bool IsNewForm = GetNewFormFlag(iPred, iParent, Dirs,
FullSet, nTerms, nPreds, nMaxTerms);
ASSERT(nTerms+1 < nMaxTerms);
UsedCols[nTerms] = UsedCols[nTerms+1] = false;
double RssBeforeKnot = RssBeforeNewTerm;
const int iNewCol = IsNewForm? nTerms+1: nTerms;
if(IsNewForm) {
// Add the new predictor as a linear term in bx[,nTerms].
// This updates RssBeforeKnot, which we will
// try to beat in FindWeightedKnot.
// Note that unlike FindPredGivenParent we never clear IsNewForm
// here (that will be done later if necessary in AddTermPair).
for(int i = 0; i < (const int)nCases; i++)
bx_(i, nTerms) = bx_(i, iParent) * x_(i, iPred);
UsedCols[nTerms] = true;
Regress(NULL, NULL, &RssBeforeKnot, NULL, NULL, NULL,
bx, yw, nCases, nResp, nMaxTerms, UsedCols);
tprintf(8,
"|Parent %-2d Pred %-2d Case -1 Cut % 12.4g< Rss %-12.5g\n",
iParent+IOFFSET, iPred+IOFFSET,
GetCut(0, iPred, nCases, x, xOrder),
MaybeZero(RssBeforeKnot));
} else {
tprintf(8, "\n|Parent %-2d Pred %-2d no new form\n",
iParent+IOFFSET, iPred+IOFFSET);
#if 1 // TODO remove this slow check when weights code has been fully tested
Regress(NULL, NULL, &RssBeforeKnot, NULL, NULL, NULL,
bx, yw, nCases, nResp, nMaxTerms, UsedCols);
if(fabs(RssBeforeKnot - RssBeforeNewTerm) > RssBeforeKnot * 1e-6)
error(
"fabs(RssBeforeKnot %g - RssBeforeNewTerm %g) %g > %g",
RssBeforeKnot, RssBeforeNewTerm,
fabs(RssBeforeKnot - RssBeforeNewTerm),
RssBeforeKnot * 1e-6);
#endif
}
double RssBestKnot = RssBeforeKnot;
int iBestCase = 0;
if(!LinPreds[iPred]) {
#if 1 // TODO remove this slow check when weights code has been fully tested
double RssTemp;
Regress(NULL, NULL, &RssTemp, NULL, NULL, NULL,
bx, yw, nCases, nResp, nMaxTerms, UsedCols);
if(fabs(RssTemp - RssBeforeKnot) > RssBeforeKnot * 1e-6)
error("fabs(RssTemp - RssBeforeKnot) > %g",
RssTemp, RssBeforeKnot,
fabs(RssTemp - RssBeforeKnot),
RssBeforeKnot * 1e-6);
#endif
ASSERT(iNewCol > 0 && iNewCol < nMaxTerms);
UsedCols[iNewCol] = true;
FindWeightedKnot(&iBestCase, &RssBestKnot,
UsedCols, iNewCol,
iParent, iPred, nCases, nResp,
bx, x, yw, xOrder, nStartSpan, nMinSpan, nEndSpan,
RssBeforeNewTerm);
}
// TODO must use NewVarAdjust here?
const bool LinPredIsBest = RssBeforeKnot <= RssBestKnot;
const double Rss = (LinPredIsBest? RssBeforeKnot: RssBestKnot);
double RssDeltaForTerm = RssBeforeNewTerm - Rss;
if(RssDeltaForTerm > *pBestRssDeltaForParent)
*pBestRssDeltaForParent = RssDeltaForTerm;
if(RssDeltaForTerm > *pBestRssDeltaForTerm) {
UpdatedBestRssDelta = true;
*pBestRssDeltaForTerm = RssDeltaForTerm;
*pLinPredIsBest = LinPredIsBest;
*piBestCase = iBestCase;
*piBestPred = iPred;
*piBestParent = iParent;
*pIsNewForm = IsNewForm;
tprintf(7,
"|Parent %-2d Pred %-2d Case %4d Cut % 12.4g Rss %-12.5g RssDelta %-12.5g%s\n",
iParent+IOFFSET,
iPred+IOFFSET,
iBestCase+IOFFSET,
GetCut(iBestCase, iPred, nCases, x, xOrder),
MaybeZero(Rss),
MaybeZero(*pBestRssDeltaForTerm),
RssDeltaForTerm > *pBestRssDeltaForTerm?
" best for term": "");
} else
tprintf(7,
"|Parent %-2d Pred %-2d Case %4d Cut % 12.4g Rss %-12.5g RssDelta %-12.5g\n",
iParent+IOFFSET,
iPred+IOFFSET,
-1,
GetCut(0, iPred, nCases, x, xOrder),
MaybeZero(Rss),
MaybeZero(RssBeforeKnot));
}
} // for iPred
if(UpdatedBestRssDelta && nUses[*piBestPred] == 0) {
// de-adjust for NewVarPenalty (only makes a difference if NewVarPenalty != 0)
const double NewVarAdjust = 1 + NewVarPenalty;
*pBestRssDeltaForTerm *= NewVarAdjust;
}
free1(UsedCols);
}
#endif // WEIGHTS
//-----------------------------------------------------------------------------
// Find a new term to add to the model, if possible, and return the
// selected case (i.e. knot), predictor, and parent term indices.
//
// The new term is a copy of an existing parent term but extended
// by multiplying the parent by a new hinge function at the selected knot.
//
// Actually, this usually finds a term _pair_, with left and right hinges.
//
// There are currently nTerms in the model. We want to add a term at index nTerms.
static void FindTerm(
int* piBestCase, // out: return -1 if no new term available, else row index
int* piBestPred, // out:
int* piBestParent, // out
double* pBestRssDeltaForTerm, // out: adding new term reduces RSS this much
// will be set to 0 if no possible new term
bool* pIsNewForm, // out
bool* pLinPredIsBest, // out: true if pred should enter linearly (no knot)
double bxOrth[], // io: column nTerms overwritten
double bxOrthCenteredT[], // io: kept in sync with bxOrth
double bxOrthMean[], // io: element nTerms overwritten
#if WEIGHTS
double bx[], // io: cols at nTerms and nTerms+1 used as scratch, will be set to 0
#else
const double bx[], // in: MARS basis matrix
#endif
const double x[], // in: nCases x nPreds, unweighted x
const double y[], // in: nCases x nResp, unweighted but scaled y
const double yw[], // in: nCases x nResp, weighted and scaled y, can be NULL
const size_t nCases, // in:
const int nResp, // in: number of cols in y
const int nPreds, // in:
const int nMaxDegree, // in:
const int nTerms, // in:
const int nMaxTerms, // in:
const double yMean[], // in: vector nResp x 1
const double RssBeforeNewTerm, // in
const double MaxLegalRssDelta, // in: FindKnot rejects any changes in Rss greater than this
const bool FullSet[], // in:
const int xOrder[], // in:
const int nDegree[], // in: degree of each term, degree of intercept is 0
const int nUses[], // in: nbr of times each predictor is used in the model
const int Dirs[], // in:
const int nFastK, // in: Fast MARS K
const double NewVarPenalty, // in: penalty for adding a new variable (default is 0)
const int LinPreds[]) // in: nPreds x 1, 1 if predictor must enter linearly
{
#if !FAST_MARS // prevent compiler warning: unused parameter
int Dummy = nFastK;
ASSERT(Dummy != -999);
#endif
#if !WEIGHTS // prevent compiler warning: unused parameter
double Dummy1 = RssBeforeNewTerm;
ASSERT(Dummy1 != -999);
#endif
tprintf(7, "\n----------------------------------------------------------"
"-------------------");
tprintf(7, "\n|FindTerm: Searching for new term %-3d "
"RssDelta 0 MaxLegalRssDelta %g\n",
nTerms+IOFFSET, MaxLegalRssDelta);
*piBestCase = -1;
*pBestRssDeltaForTerm = 0;
*pLinPredIsBest = false;
*pIsNewForm = false;
int i;
xbx = (double*)malloc1(nCases * sizeof(double),
"xbx\t\t\tnCases %d sizeof(double) %d",
(const int)nCases, sizeof(double));
CovSx = (double*)malloc1(nMaxTerms * sizeof(double),
"CovSx\t\t\tnMaxTerms %d sizeof(double) %d",
nMaxTerms, sizeof(double));
CovCol = (double*)calloc1(nMaxTerms, sizeof(double),
"CovCol\t\tnMaxTerms %d sizeof(double) %d",
nMaxTerms, sizeof(double));
ycboSum = (double*)calloc1(nMaxTerms * nResp, sizeof(double),
"ycboSum\t\tnMaxTerms %d nResp %d sizeof(double) %d",
nMaxTerms, nResp, sizeof(double));
for(int iResp = 0; iResp < nResp; iResp++)
for(int iTerm = 0; iTerm < nTerms; iTerm++)
for(i = 0; i < (const int)nCases; i++)
ycboSum_(iTerm,iResp) +=
(y_(i,iResp) - yMean[iResp]) * bxOrth_(i,iTerm);
#if USING_R
const int nServiceR = (int)1e6 / nCases;
#endif
int iParent;
#if FAST_MARS
GetNextParent(true, nFastK); // init queue iterator
while((iParent = GetNextParent(false, nFastK)) > -1) {
#else
for(iParent = 0; iParent < nTerms; iParent++) {
#endif
#if USING_R
static int iServiceR = 0;
if(++iServiceR > nServiceR) {
ServiceR();
iServiceR = 0;
}
#endif
// Assume a bad RssDelta for iParent. This pushes parent terms that
// can't be used to the bottom of the FastMARS queue. (A parent can't
// be used if its degree is too big or all predictors are in the parent.)
double BestRssDeltaForParent = -1; // used only by FAST_MARS
if(nDegree[iParent] >= nMaxDegree)
tprintf(7, "|Parent %-2d %52.52s skip (degree of term would be %d)\n",
iParent+IOFFSET, " ", nDegree[iParent]+1);
else {
int nMinSpan, nEndSpan, nStartSpan;
GetSpanParams(&nMinSpan, &nEndSpan, &nStartSpan,
nCases, nPreds, nDegree[iParent]+1, iParent, bx);
#if WEIGHTS
if(yw)
FindWeightedPredGivenParent(
piBestCase, piBestPred, piBestParent, pBestRssDeltaForTerm,
&BestRssDeltaForParent, pIsNewForm, pLinPredIsBest,
bx, RssBeforeNewTerm,
iParent, x, yw, nCases, nResp, nPreds, nTerms, nMaxTerms,
FullSet, xOrder, nUses, Dirs, NewVarPenalty, LinPreds,
nMinSpan, nEndSpan, nStartSpan);
else
#endif
FindPredGivenParent(
piBestCase, piBestPred, piBestParent, pBestRssDeltaForTerm,
&BestRssDeltaForParent, pIsNewForm, pLinPredIsBest,
bxOrth, bxOrthCenteredT, bxOrthMean, xbx, CovSx, CovCol, ycboSum,
bx, yMean, RssBeforeNewTerm, MaxLegalRssDelta,
iParent, x, y, nCases, nResp, nPreds, nTerms, nMaxTerms,
FullSet, xOrder, nUses, Dirs, NewVarPenalty, LinPreds,
nMinSpan, nEndSpan, nStartSpan);
#if FAST_MARS
UpdateRssDeltaInQ(iParent, nTerms, BestRssDeltaForParent);
#endif
}
} // iParent
tprintf(7, "\n");
// free in opposite order to alloc to help operating system memory manager
free1(ycboSum);
free1(CovCol);
free1(CovSx);
free1(xbx);
}
//-----------------------------------------------------------------------------
static void PrintForwardProlog(
const size_t nCases, // in
const int nPreds, // in
const int nMaxTerms, // in
const char* sPredNames[], // in: predictor names, can be NULL
const bool HasWeights) // in
{
if(TraceGlobal == 1)
printf("Forward pass term %d", IOFFSET);
else if(TraceGlobal == 1.5)
printf("Forward pass term %d\n", IOFFSET);
else if(TraceGlobal >= 2) {
int nMinSpan, nEndSpan, nStartSpan;
GetSpanParams(&nMinSpan, &nEndSpan, &nStartSpan,
nCases, nPreds, 1 /*nDegree*/, 0 /*iParent*/, NULL /*bx*/);
char sx[100];
strcpy(sx, sFormatMemSize(nCases * nPreds * sizeof(double), false));
char sbx[100];
strcpy(sbx, sFormatMemSize(nCases * nMaxTerms * sizeof(double), false));
printf("Forward pass: minspan %d endspan %d x[%d,%d] %s bx[%d,%d] %s%s\n\n",
nMinSpan, nEndSpan,
(int)nCases, nPreds, sx,
(int)nCases, nMaxTerms, sbx,
HasWeights? " weighted": "");
printf(" GRSq RSq DeltaRSq Pred ");
if(sPredNames)
printf(" PredName ");
printf(" Cut Terms Par Deg\n");
// following matches printfs in PrintForwardStep
if(sPredNames) // in: predictor names, can be NULL
printf("%-4d%9.4f %6.4f %12.12s\n",
IOFFSET, 0., 0., "(Intercept)");
else
printf("%-4d%9.4f %6.4f %d\n",
IOFFSET, 0., 0., IOFFSET);
}
}
//-----------------------------------------------------------------------------
static void PrintForwardStep(
const int nTerms,
const int nUsedTerms,
const int iBestCase,
const int iBestPred,
const int iBestParent,
const int nDegree,
const double RSq,
const double RSqDelta,
const double Gcv,
const double GcvNull,
const size_t nCases,
const int xOrder[],
const double x[],
const bool IsTermPair,
const char* sPredNames[]) // in: predictor names, can be NULL
{
if(TraceGlobal == 6)
printf("\n\n");
if(TraceGlobal == 1) {
printf(", ");
if(nTerms % 30 == 29)
printf("\n ");
printf("%d", nTerms+IOFFSET);
} else if(TraceGlobal == 1.5)
printf("Forward pass term %d\n", nTerms+IOFFSET);
else if(TraceGlobal >= 2) {
tprintf(7,
" GRSq RSq DeltaRSq Pred PredName Cut Terms Par Deg\n");
printf("%-4d%9.4f %6.4f %12.4g ",
nTerms+IOFFSET, 1 - Gcv / GcvNull, RSq, RSqDelta);
if(iBestPred < 0) // *piBestCase not updated in FindKnot (no DeltaRSq)
printf(" - ");
else {
printf("%4d", iBestPred+IOFFSET);
if(sPredNames) {
if(sPredNames[iBestPred] && sPredNames[iBestPred][0])
printf(" %12.12s ", sPredNames[iBestPred]);
else
printf(" %12.12s ", " ");
}
if(iBestCase == -1)
printf(" none ");
else
printf("% 11.5g%c ",
GetCut(iBestCase, iBestPred, nCases, x, xOrder),
(iBestCase==0? '<': ' ')); // print '<' if knot is at min
if(IsTermPair) // two new used terms?
printf("%-3d %-3d ", nUsedTerms-2+IOFFSET, nUsedTerms-1+IOFFSET);
else
printf("%-3d ", nUsedTerms-1+IOFFSET);
if(iBestParent != 0) // print parent if it isn't the intercept
tprintf(2, "%3d ", iBestParent+IOFFSET);
else
tprintf(2, " ", iBestParent+IOFFSET);
printf("%3d ", nDegree);
}
}
#if !USING_R // no flush needed when using R_printf
if(TraceGlobal != 0)
fflush(stdout);
#endif
}
//-----------------------------------------------------------------------------
static int ForwardEpilog( // returns reason we stopped adding terms
const int nTerms, const int nMaxTerms,
const double Thresh,
const double RSq, const double RSqDelta,
const double Gcv, const double GcvNull,
const int iBestCase,
const bool FullSet[])
{
tprintf(7,
"\n-----------------------------------------------------------------------------\n");
const double GRSq = 1 - Gcv / GcvNull;
int iTermCond = 0;
char sUsed[100] = "";
const int nUsed = GetNbrUsedCols(FullSet, nMaxTerms);
if(nUsed != nTerms)
sprintf(sUsed, ", %d term%s used", nUsed, nUsed == 1? "": "s");
char sTerms[200]; // May 2018: changed 100 to 200 for specious CRAN warning: '%s' directive writing up to 99 bytes into a region of size between 84 and 94 [-Wformat-overflow=]
sprintf(sTerms, "%d term%s%s", nTerms, nTerms == 1? "": "s", sUsed);
// NOTE 1: this code must match the loop termination conditions in ForwardPass
// NOTE 2: if you update this, also update print.termcond in the R code
// treat very low nMaxTerms as a special case
// because RSDelta etc. not yet completely initialized
if(nMaxTerms < 3) {
iTermCond = 1;
tprintf(1, "\nReached maximum number of terms %d\n", nMaxTerms);
} else if(Thresh != 0 && GRSq < MIN_GRSQ) {
if(GRSq < -1000) {
iTermCond = 2;
tprintf(1, "\nGRSq -Inf at %s\n", sTerms);
} else {
iTermCond = 3;
if(TraceGlobal >= 1)
printf("\nReached minimum GRSq %g at %s (GRSq %.2g)\n",
MIN_GRSQ, sTerms, GRSq);
}
} else if(Thresh != 0 && RSqDelta < Thresh) {
iTermCond = 4;
if(TraceGlobal >= 1)
printf("\nRSq changed by less than %g at %s (DeltaRSq %.2g)\n",
Thresh, sTerms, RSqDelta);
} else if(RSq >= 1-Thresh) {
iTermCond = 5;
if(TraceGlobal >= 1)
printf("\nReached maximum RSq %.4f at %s (RSq %.4f)\n",
1-Thresh, sTerms, RSq);
} else if(iBestCase < 0) { // TODO seems fishy, happens with linpreds so should give appropriate msg?
iTermCond = 6;
tprintf(1,
"\nNo new term increases RSq (perhaps reached numerical limits) at %s\n",
sTerms);
} else {
iTermCond = 7;
#if USING_R
tprintf(1, "\nReached nk %d\n", nMaxTerms);
#else
tprintf(1, "\nReached maximum number of terms %d\n", nMaxTerms);
#endif
}
if(TraceGlobal >= 1)
printf("After forward pass GRSq %.3f RSq %.3f\n", GRSq, RSq);
tprintf(2, "Forward pass complete: %s\n", sTerms);
tprintf(3, "\n");
return iTermCond;
}
//-----------------------------------------------------------------------------
static void CheckVec(
const double x[],
const size_t nCases,
const int nCols,
const char sVecName[])
{
int iCol, i;
for(iCol = 0; iCol < nCols; iCol++)
for(i = 0; i < (const int)nCases; i++) {
#if USING_R
if(ISNA(x[i + iCol * nCases])) {
if(nCols > 1)
error("%s[%d,%d] is NA",
sVecName, i+IOFFSET, iCol+IOFFSET);
else
error("%s[%d] is NA", sVecName, i+IOFFSET);
}
#endif
if(ISNAN(x[i + iCol * nCases])) {
if(nCols > 1)
error("%s[%d,%d] is NaN",
sVecName, i+IOFFSET, iCol+IOFFSET);
else
error("%s[%d] is NaN", sVecName, i+IOFFSET);
}
if(!FINITE(x[i + iCol * nCases])) {
if(nCols > 1)
error("%s[%d,%d] is not finite",
sVecName, i+IOFFSET, iCol+IOFFSET);
else
error("%s[%d] is not finite", sVecName, i+IOFFSET);
}
}
}
//-----------------------------------------------------------------------------
static double CheckRssNull(
double RssNull,
const double y[],
const int iResp,
const int nResp,
const size_t nCases)
{
if(RssNull < 1e-8 * nCases) { // 1e-8 is arbitrary
if(nResp)
tprintf(1, "Variance of y[,%d] is zero (values are all equal to %g)\n",
iResp+IOFFSET, y_(0,iResp));
else
tprintf(1, "Variance of y is zero (values are all equal to %g)\n",
y_(0,iResp));
RssNull = 1e-8 * nCases; // prevent later divide by zero
}
return RssNull;
}
//-----------------------------------------------------------------------------
static double GetRssNull(
const double y[], // in: nCases x nResp, unweighted but scaled y
const double WeightsArg[], // in: nCases x 1, can be NULL
const size_t nCases, // in: number of rows in x and elements in y
const int nResp) // in: number of cols in y
{
double RssNull = 0;
if(WeightsArg)
for(int iResp = 0; iResp < nResp; iResp++) {
double SumY = 0, SumWeights = 0;
int i;
for(i = 0; i < (const int)nCases; i++) {
SumY += WeightsArg[i] * y_(i,iResp);
SumWeights += WeightsArg[i];
}
const double WeightedMean = SumY / SumWeights;
for(i = 0; i < (const int)nCases; i++) {
RssNull += WeightsArg[i] * sq(y_(i,iResp) - WeightedMean);
}
RssNull = CheckRssNull(RssNull, y, iResp, nResp, nCases);
}
else // no weights
for(int iResp = 0; iResp < nResp; iResp++) {
const double yMean = Mean(&y_(0,iResp), nCases);
RssNull += SumOfSquares(&y_(0,iResp), yMean, nCases);
RssNull = CheckRssNull(RssNull, y, iResp, nResp, nCases);
}
return RssNull;
}
//-----------------------------------------------------------------------------
// The limits below are somewhat arbitrary and generous. They are intended to
// catch gross errors on the part of the caller, and to prevent crashes because
// of 0 sizes etc. We use error rather than ASSERT because these are user
// settable params and we want to be informative from the user's perspective.
// The errors are reported using the variable names in the R code.
static void CheckForwardPassArgs(
const double x[],
const double y[],
const double yw[],
const double WeightsArg[],
const size_t nCases,
const int nResp,
const int nPreds,
const int nMaxDegree,
const int nMaxTerms,
const double Penalty,
const double Thresh,
const double FastBeta,
const double NewVarPenalty,
const int LinPreds[],
const double AdjustEndSpan,
const bool AutoLinPreds,
const bool UseBetaCache)
{
const int nCases1 = (const int)nCases; // type convert from size_t
if(nCases1 < 2)
error("the x matrix must have at least two rows");
if(nCases1 > 1e9)
error("too many rows %d in the input matrix, max allowed is 1e9", nCases1);
if(nResp < 1)
error("the number of responses %d is less than 1", nResp);
if(nResp > 1000)
error("the number of responses %d is greater than 1000", nResp);
if(nPreds < 1)
error("the number of predictors %d is less than 1", nPreds);
if(nPreds > 1e5)
error("the number of predictors %d is greater than 1e5", nPreds);
if(nMaxDegree <= 0)
error("degree %d is not greater than 0", nMaxDegree);
if(nMaxDegree > MAX_DEGREE)
error("degree %d is greater than %d", nMaxDegree, MAX_DEGREE);
if(nMaxTerms < 1) // prevent internal misbehavior
error("nk %d is less than 1", nMaxTerms);
if(nMaxTerms > 1000)
error("nk %d is greater than 1000", nMaxTerms);
if(Penalty < 0 && Penalty != -1)
error("penalty %g is less than 0 and the only legal value less than 0 is -1 "
"(meaning terms and knots are free)", Penalty);
if(Penalty > 1000)
error("penalty %g is greater than 1000", Penalty);
if(Thresh < 0)
error("thresh %g is less than 0", Thresh);
if(Thresh >= 1)
error("thresh %g >= 1", Thresh);
if(nMinSpanGlobal > nCases1)
error("minspan %d is greater than the number of cases %d", nMinSpanGlobal, nCases1);
if(nEndSpanGlobal > nCases1)
error("endspan %d is greater than the number of cases %d", nEndSpanGlobal, nCases1);
else if(nEndSpanGlobal < 0)
error("endspan %d is less than 0", nEndSpanGlobal);
if(FastBeta < 0)
error("fast.beta %g is less than 0", FastBeta);
if(FastBeta > 1000)
error("fast.beta %g is greater than 1000", FastBeta);
if(TraceGlobal < 0)
warning("trace %g is less than 0", TraceGlobal);
if(TraceGlobal > 10)
warning("trace %g is greater than 10", TraceGlobal);
if(NewVarPenalty < 0)
warning("newvar.penalty %g is less than 0", NewVarPenalty);
if(NewVarPenalty > 100)
warning("newvar.penalty %g is greater than 100", NewVarPenalty);
if(AdjustEndSpan < 0 || AdjustEndSpan > 10)
error("Endspan.penalty is %g but should be between 0 and 10", AdjustEndSpan);
if(AutoLinPreds != 0 && AutoLinPreds != 1)
error("Auto.linpreds is neither TRUE nor FALSE");
if(UseBetaCache != 0 && UseBetaCache != 1)
warning("Use.Beta.Cache is neither TRUE nor FALSE");
CheckVec(x, nCases, nPreds, "x");
CheckVec(y, nCases, nResp, "y");
#if WEIGHTS
if(yw) {
ASSERT(WeightsArg);
CheckVec(yw, nCases, nResp, "yw");
for(int i = 0; i < (const int)nCases; i++) {
CheckVec(WeightsArg, nCases, 1, "weights");
if(WeightsArg[i] < ALMOST_ZERO)
error("weights[%d] is not greater than zero",
i+IOFFSET);
}
}
#else
ASSERT(!yw);
ASSERT(!WeightsArg || abs(WeightsArg[0] - WeightsArg[1]) < 1e-8);
#endif
for(int iPred = 0; iPred < nPreds; iPred++)
if(LinPreds[iPred] != 0 && LinPreds[iPred] != 1)
error("linpreds[%d] is not 0 or 1", iPred+IOFFSET);
}
//-----------------------------------------------------------------------------
// Forward pass
//
// After initializing the intercept term, the main for loop adds terms in pairs.
// In the for loop, nTerms is the index of the potential new term; nTerms+1
// the index of its partner.
// The upper term in the term pair may not be useable. If so we still
// increment nTerms by 2 but don't set the flag in FullSet.
static void ForwardPass(
int* pnTerms, // out: highest used term number in full model
int* piTermCond, // out: reason we terminated the forward pass
bool FullSet[], // out: 1 * nMaxTerms, indices of lin indep cols of bx
double bx[], // out: MARS basis matrix, nCases * nMaxTerms
int Dirs[], // out: nMaxTerms * nPreds, -1,0,1,2 for iTerm, iPred
double Cuts[], // out: nMaxTerms * nPreds, cut for iTerm, iPred
int nDegree[], // out: degree of each term, degree of intercept is 0
int nUses[], // out: nbr of times each predictor is used in the model
const double x[], // in: nCases x nPreds, unweighted x
const double y[], // in: nCases x nResp, unweighted but scaled y
const double yw[], // in: nCases x nResp, weighted and scaled y, can be NULL
const double WeightsArg[], // in: nCases x 1, can be NULL
const size_t nCases, // in: number of rows in x and elements in y
const int nResp, // in: number of cols in y
const int nPreds, // in:
const int nMaxDegree, // in:
const int nMaxTerms, // in:
const double Penalty, // in: GCV penalty per knot
const double Thresh, // in: forward step threshold
int nFastK, // in: Fast MARS K
const double FastBeta, // in: Fast MARS ageing coef
const double NewVarPenalty, // in: penalty for adding a new variable (default is 0)
const int LinPreds[], // in: nPreds x 1, 1 if predictor must enter linearly
const double AdjustEndSpan, // in:
const bool AutoLinPreds, // in: assume predictor linear if knot is min predictor value
const bool UseBetaCache, // in: true to use the beta cache, for speed
const char* sPredNames[]) // in: predictor names, can be NULL
{
tprintf(5, "earth.c %s\n", VERSION);
CheckForwardPassArgs(x, y, yw, WeightsArg, nCases, nResp, nPreds,
nMaxDegree, nMaxTerms, Penalty, Thresh, FastBeta, NewVarPenalty,
LinPreds, AdjustEndSpan, AutoLinPreds, UseBetaCache);
if(nFastK <= 0)
nFastK = 10000+1; // bigger than any nMaxTerms
if(nFastK < 3) // avoid possible queue boundary conditions
nFastK = 3;
xOrder = GetArrayOrder(x, nCases, nPreds);
InitBetaCache(UseBetaCache, nMaxTerms, nPreds);
bxOrth = (double*)malloc1(nCases * nMaxTerms * sizeof(double),
"bxOrth\t\tnCases %d nMaxTerms %d sizeof(double) %d",
(const int)nCases, nMaxTerms, sizeof(double));
bxOrthCenteredT = (double*)malloc1(nMaxTerms * nCases * sizeof(double),
"bxOrthCenteredT\tnMaxTerms %d nCases %d sizeof(double) %d",
nMaxTerms, (const int)nCases, sizeof(double));
bxOrthMean = (double*)malloc1(nMaxTerms * nResp * sizeof(double),
"bxOrthMean\t\tnMaxTerms %d nResp %d sizeof(double) %d",
nMaxTerms, nResp, sizeof(double));
yMean = (double*)malloc1(nResp * sizeof(double),
"yMean\t\t\tnResp %d sizeof(double) %d",
nResp, sizeof(double));
memset(FullSet, 0, nMaxTerms * sizeof(bool));
memset(Dirs, 0, nMaxTerms * nPreds * sizeof(int));
memset(Cuts, 0, nMaxTerms * nPreds * sizeof(double));
memset(nDegree, 0, nMaxTerms * sizeof(int));
memset(nUses, 0, nPreds * sizeof(int));
memset(bx, 0, nCases * nMaxTerms * sizeof(double));
// Intercept columns of bx and bxOrth.
// Note that we use the weights here, and since every term is a multiple
// of this intercept term or a term derived from it, we don't need to use
// the weighted x when forming bx.
if(WeightsArg)
for(int i = 0; i < (const int)nCases; i++)
bx_(i,0) = sqrt(WeightsArg[i]);
else
for(int i = 0; i < (const int)nCases; i++)
bx_(i,0) = 1;
bool GoodCol;
InitBxOrthCol(bxOrth, bxOrthCenteredT, bxOrthMean, &GoodCol,
&bx_(0,0), 0 /*nTerms*/, FullSet, nCases, nMaxTerms, -1, -1);
if(!GoodCol) // should never happen
tprintf(1, "GoodCol is false in ForwardPass\n");
GoodCol = true;
FullSet[0] = true; // intercept
for(int iResp = 0; iResp < nResp; iResp++)
yMean[iResp] = Mean(&y_(0,iResp), nCases);
const double RssNull = GetRssNull(y, WeightsArg, nCases, nResp);
double Rss = RssNull, RssDelta = RssNull, RSq = 0, RSqDelta = 0;
int nUsedTerms = 1; // number of used basis terms including intercept, for GCV calc
double Gcv = 0, GcvNull = GetGcv(nUsedTerms, nCases, RssNull, Penalty);
PrintForwardProlog(nCases, nPreds, nMaxTerms, sPredNames, yw != NULL);
#if FAST_MARS
InitQ(nMaxTerms);
AddTermToQ(0, 1, RssNull, true, nMaxTerms, FastBeta); // intercept term into Q
#endif
int nTerms = 1, iBestCase = -1;
if(nMaxTerms >= 3) while(1) { // start after intercept, add terms in pairs
int iBestPred = -1, iBestParent = -1;
bool IsNewForm, LinPredIsBest;
#if USING_R
ServiceR();
#endif
if(Rss <= 0)
error("assertion failed: Rss <= 0 (y is all const?)");
ASSERT(RssDelta > 0);
// Changed factor from 2 to 10 in version 4.2.0 (2 was too conservative).
// Note that only the code without weights uses this.
const double MaxLegalRssDelta = min(1.01 * Rss, 10 * RssDelta);
FindTerm(&iBestCase, &iBestPred, &iBestParent,
&RssDelta, &IsNewForm, &LinPredIsBest,
bxOrth, bxOrthCenteredT, bxOrthMean,
bx, x, y, yw, nCases, nResp, nPreds, nMaxDegree, nTerms, nMaxTerms,
yMean, Rss, MaxLegalRssDelta, FullSet, xOrder, nDegree,
nUses, Dirs, nFastK, NewVarPenalty, LinPreds);
// following code added for Auto.linpreds (earth version 4.6.0, Dec 2017)
if((LinPredIsBest && iBestCase != 0) || // paranoia, should never happen
(!LinPredIsBest && iBestCase == 0))
printf("\nLinPredIsBest %d yet iBestCase%-5d\n", LinPredIsBest, iBestCase);
if(!AutoLinPreds && !LinPreds[iBestPred])
LinPredIsBest = false;
if(iBestCase >= 0)
AddTermPair(Dirs, Cuts, bx, bxOrth, bxOrthCenteredT, bxOrthMean,
FullSet, &IsNewForm, nDegree, nUses,
nTerms, iBestParent, iBestCase, iBestPred, nPreds, nCases,
nMaxTerms, LinPredIsBest, LinPreds, x, xOrder, yw != NULL);
const bool IsTermPair = iBestCase > 0 && IsNewForm;
nUsedTerms++;
if(IsTermPair)
nUsedTerms++; // add paired term
Rss -= RssDelta;
Rss = MaybeZero(Rss); // RSS can go slightly neg due to numerical error
Gcv = GetGcv(nUsedTerms, nCases, Rss, Penalty);
const double OldRSq = RSq;
RSq = 1 - Rss / RssNull;
RSqDelta = MaybeZero(RSq - OldRSq);
PrintForwardStep(nTerms, nUsedTerms, iBestCase, iBestPred,
iBestParent, iBestParent < 0? 0: nDegree[iBestParent]+1,
RSq, RSqDelta, Gcv, GcvNull,
nCases, xOrder, x, IsTermPair, sPredNames);
// note the possible breaks in the code below
if(iBestCase < 0) { // *piBestCase was not updated in FindKnot
FullSet[nTerms] = FullSet[nTerms+1] = false;
tprintf(2, "reject (no DeltaRsq)\n");
break;
}
if(Thresh != 0 && RSqDelta < Thresh) {
FullSet[nTerms] = FullSet[nTerms+1] = false;
tprintf(2, "reject (small DeltaRSq)\n");
break;
}
if(Thresh != 0 && 1 - Gcv / GcvNull < MIN_GRSQ) {
FullSet[nTerms] = FullSet[nTerms+1] = false;
tprintf(2, "reject (negative GRSq)\n");
break;
}
#if FAST_MARS
if(!LinPredIsBest && IsNewForm) { // good upper term?
AddTermToQ(nTerms, nTerms, POS_INF, false, nMaxTerms, FastBeta);
AddTermToQ(nTerms+1, nTerms, POS_INF, true, nMaxTerms, FastBeta);
} else
AddTermToQ(nTerms, nTerms, POS_INF, true, nMaxTerms, FastBeta);
if(TraceGlobal == 6)
PrintSortedQ(nFastK);
#endif
nTerms += 2;
if(RSq >= 1 - Thresh) {
tprintf(2, "final (max RSq)\n");
break;
}
if(nTerms >= nMaxTerms - 1) { // -1 allows for upper term in pair
tprintf(2, "final (reached nk %d)\n", nMaxTerms);
break;
}
tprintf(2, "\n");
} // while(1)
*piTermCond = ForwardEpilog(nTerms, nMaxTerms, Thresh, RSq, RSqDelta,
Gcv, GcvNull, iBestCase, FullSet);
*pnTerms = nTerms;
// free in opposite order to alloc to help operating system memory manager
#if FAST_MARS
FreeQ();
#endif
free1(yMean);
free1(bxOrthMean);
free1(bxOrthCenteredT);
free1(bxOrth);
FreeBetaCache();
free1(xOrder);
}
//-----------------------------------------------------------------------------
// This is an interface from R to the C routine ForwardPass
#if USING_R
SEXP ForwardPassR( // for use by R
SEXP SEXP_FullSet, // out: nMaxTerms x 1, bool vec of lin indep cols of bx
SEXP SEXP_bx, // out: MARS basis matrix, nCases x nMaxTerms
SEXP SEXP_Dirs, // out: nMaxTerms x nPreds, elements are -1,0,1,2
SEXP SEXP_Cuts, // out: nMaxTerms x nPreds, cut for iTerm,iPred
SEXP SEXP_iTermCond, // out: reason we terminated the forward pass
SEXP SEXP_x, // in: nCases x nPreds, unweighted x
SEXP SEXP_y, // in: nCases x nResp, unweighted but scaled y
SEXP SEXP_yw, // in: nCases x nResp, weighted and scaled y
SEXP SEXP_WeightsArg, // in: nCases x 1, never R_NilValue
SEXP SEXP_nCases, // in: number of rows in x and elements in y
SEXP SEXP_nResp, // in: number of cols in y
SEXP SEXP_nPreds, // in: number of cols in x
SEXP SEXP_nMaxDegree, // in:
SEXP SEXP_Penalty, // in:
SEXP SEXP_nMaxTerms, // in:
SEXP SEXP_Thresh, // in: forward step threshold
SEXP SEXP_nMinSpan, // in:
SEXP SEXP_nEndSpan, // in:
SEXP SEXP_nFastK, // in: Fast MARS K
SEXP SEXP_FastBeta, // in: Fast MARS ageing coef
SEXP SEXP_NewVarPenalty, // in: penalty for adding a new variable (default is 0)
SEXP SEXP_LinPreds, // in: nPreds x 1, 1 if predictor must enter linearly
SEXP SEXP_Allowed, // in: constraints function, can be MyNullFunc
SEXP SEXP_nAllowedArgs, // in: number of arguments to Allowed function, 3 or 4
SEXP SEXP_Env, // in: environment for Allowed function
SEXP SEXP_AdjustEndSpan, // in:
SEXP SEXP_nAutoLinPreds, // in: assume predictor linear if knot is min predictor value
SEXP SEXP_nUseBetaCache, // in: 1 to use the beta cache, for speed
SEXP SEXP_Trace, // in: 0 none 1 overview 2 forward 3 pruning 4 more pruning
SEXP SEXP_sPredNames) // in: predictor names in trace printfs
{
const size_t nCases = (size_t)(INTEGER(SEXP_nCases)[0]);
const int nResp = INTEGER(SEXP_nResp)[0];
const int nPreds = INTEGER(SEXP_nPreds)[0];
const int nMaxTerms = INTEGER(SEXP_nMaxTerms)[0];
nMinSpanGlobal = INTEGER(SEXP_nMinSpan)[0];
nEndSpanGlobal = INTEGER(SEXP_nEndSpan)[0];
AdjustEndSpanGlobal = REAL(SEXP_AdjustEndSpan)[0];
TraceGlobal = REAL(SEXP_Trace)[0];
// nUses is the number of time each predictor is used in the model
nUses = (int*)malloc1(nPreds * sizeof(int),
"nUses\t\t\tnPreds %d sizeof(int) %d",
nPreds, sizeof(int));
// nDegree is degree of each term, degree of intercept is considered to be 0
nDegree = (int*)malloc1(nMaxTerms * sizeof(int),
"nDegree\t\tnMaxTerms %d sizeof(int) %d",
nMaxTerms, sizeof(int));
iDirs = (int*)calloc1(nMaxTerms * nPreds, sizeof(int),
"iDirs\t\t\tnMaxTerms %d nPreds %d sizeof(int) %d",
nMaxTerms, nPreds, sizeof(int));
// convert FullSet int to bool (may be redundant, depending on compiler)
BoolFullSet = (bool*)malloc1(nMaxTerms * sizeof(bool),
"BoolFullSet\t\tnMaxTerms %d sizeof(bool) %d",
nMaxTerms, sizeof(bool));
int iTerm;
for(iTerm = 0; iTerm < nMaxTerms; iTerm++)
BoolFullSet[iTerm] = INTEGER(SEXP_FullSet)[iTerm] != 0;
// copy predictor names from SEXP_sPredNames to sPredNames
ASSERT(LENGTH(SEXP_sPredNames) == nPreds);
const char** sPredNames = (const char**)malloc1(
LENGTH(SEXP_sPredNames) * sizeof(char*),
"sPredNames\t\tLENGTH(SEXP_sPredNames) %d sizeof(char*) %d",
nPreds, sizeof(char*));
for(int i = 0; i < nPreds; i++)
sPredNames[i] = (char*)CHAR(STRING_ELT(SEXP_sPredNames, i));
#if !WEIGHTS
ASSERT(SEXP_yw == R_NilValue);
#endif
ASSERT(SEXP_WeightsArg != R_NilValue);
InitAllowedFunc(SEXP_Allowed, INTEGER(SEXP_nAllowedArgs)[0], SEXP_Env,
sPredNames, nPreds);
int nTerms;
ForwardPass(&nTerms, INTEGER(SEXP_iTermCond),
BoolFullSet, REAL(SEXP_bx), iDirs, REAL(SEXP_Cuts), nDegree, nUses,
REAL(SEXP_x), REAL(SEXP_y),
(SEXP_yw == R_NilValue)? NULL: REAL(SEXP_yw),
REAL(SEXP_WeightsArg), nCases, nResp, nPreds,
INTEGER(SEXP_nMaxDegree)[0], nMaxTerms, REAL(SEXP_Penalty)[0],
REAL(SEXP_Thresh)[0], INTEGER(SEXP_nFastK)[0],
REAL(SEXP_FastBeta)[0], REAL(SEXP_NewVarPenalty)[0],
INTEGER(SEXP_LinPreds), REAL(SEXP_AdjustEndSpan)[0],
INTEGER(SEXP_nAutoLinPreds)[0], INTEGER(SEXP_nUseBetaCache)[0] != 0,
sPredNames);
FreeAllowedFunc();
// remove linearly independent columns if necessary -- this updates BoolFullSet
RegressAndFix(NULL, NULL, NULL, BoolFullSet,
REAL(SEXP_bx),
(SEXP_yw == R_NilValue)? REAL(SEXP_y): REAL(SEXP_yw),
nCases, nResp, nMaxTerms);
double* p = REAL(SEXP_Dirs);
for(iTerm = 0; iTerm < nMaxTerms; iTerm++) // convert int to double
for(int iPred = 0; iPred < nPreds; iPred++)
p[iTerm + iPred * nMaxTerms] = iDirs[iTerm + iPred * nMaxTerms];
for(iTerm = 0; iTerm < nMaxTerms; iTerm++) // convert bool to int
INTEGER(SEXP_FullSet)[iTerm] = BoolFullSet[iTerm];
free1(sPredNames);
free1(BoolFullSet);
free1(iDirs);
free1(nDegree);
free1(nUses);
return R_NilValue;
}
#endif // USING_R
//-----------------------------------------------------------------------------
// Step backwards through the terms, at each step deleting the term that
// causes the least RSS increase. The subset of terms and RSS of each subset are
// saved in PruneTerms and RssVec (which are indexed on subset size).
//
// The crux of the method used here is that the change in RSS (for nResp=1)
// caused by removing predictor iPred is DeltaRss = sq(Betas[iPred]) / Diags[iPred]
// where Diags is the diagonal elements of the inverse of X'X.
// See for example Miller (see refs in file header) section 3.4 p44.
//
// For multiple responses we sum the above DeltaRss over all responses.
//
// This method is fast and simple but accuracy can be poor if inv(X'X) is
// ill conditioned. The Alan Miller code in the R package "leaps" uses a more
// stable method, but does not support multiple responses.
//
// The "Xtx" in the name refers to the X'X matrix.
static void EvalSubsetsUsingXtx(
bool PruneTerms[], // out: nMaxTerms x nMaxTerms
double RssVec[], // out: nMaxTerms x 1, RSS of each subset
const size_t nCases, // in
const int nResp, // in: number of cols in y
const int nMaxTerms, // in: number of MARS terms in full model
const double bx[], // in: nCases x nMaxTerms, all cols must be indep
const double y[]) // in: nCases * nResp
{
double* Betas = (double*)malloc1(nMaxTerms * nResp * sizeof(double),
"Betas\t\t\tnMaxTerms %d nResp %d sizeof(double) %d",
nMaxTerms, nResp, sizeof(double));
double* Diags = (double*)malloc1(nMaxTerms * sizeof(double),
"Diags\t\t\tnMaxTerms %d sizeof(double) %d",
nMaxTerms, sizeof(double));
WorkingSet = (bool*)malloc1(nMaxTerms * sizeof(bool),
"WorkingSet\t\tnMaxTerms %d sizeof(bool) %d",
nMaxTerms, sizeof(bool));
for(int iTerm = 0; iTerm < nMaxTerms; iTerm++)
WorkingSet[iTerm] = true;
const double RssNull = GetRssNull(y, NULL, nCases, nResp);
tprintf(4, "EvalSubsetsUsingXtx:\nnTerms iTerm DeltaRss RSq\n");
for(int nUsedCols = nMaxTerms; nUsedCols > 0; nUsedCols--) {
int nRank;
double Rss;
Regress(Betas, NULL, &Rss, Diags, &nRank, NULL,
bx, y, nCases, nResp, nMaxTerms, WorkingSet);
if(nRank != nUsedCols)
error("nRank %d != nUsedCols %d "
"(probably because of lin dep terms in bx)\n",
nRank, nUsedCols);
RssVec[nUsedCols-1] = Rss;
memcpy(PruneTerms + (nUsedCols-1) * nMaxTerms, WorkingSet,
nMaxTerms * sizeof(bool));
if(nUsedCols == 1)
break;
// set iDelete to the best term for deletion
int iDelete = -1; // term to be deleted
int iTerm1 = 0; // index taking into account false vals in WorkingSet
double MinDeltaRss = POS_INF;
for(int iTerm = 0; iTerm < nMaxTerms; iTerm++) {
if(WorkingSet[iTerm]) {
double DeltaRss = 0;
for(int iResp = 0; iResp < nResp; iResp++)
DeltaRss += sq(Betas_(iTerm1, iResp)) / Diags[iTerm1];
bool NewMin = false;
if(iTerm > 0 && DeltaRss < MinDeltaRss) { // new minimum?
MinDeltaRss = DeltaRss;
iDelete = iTerm;
NewMin = true;
}
if(iTerm != 0)
tprintf(4, "%6d %5d %8.5g %6.4f%s\n",
nUsedCols, iTerm+IOFFSET, DeltaRss,
1 - (Rss + DeltaRss) / RssNull, NewMin? " min" : "");
iTerm1++;
}
}
ASSERT(iDelete > 0);
WorkingSet[iDelete] = false;
tprintf(4, "\n");
}
free1(WorkingSet);
free1(Diags);
free1(Betas);
}
//-----------------------------------------------------------------------------
// This is invoked from R if y has multiple columns i.e. a multiple response model.
// It is needed because the alternative (Alan Miller's Fortran code) supports
// only one response.
#if USING_R
void EvalSubsetsUsingXtxR( // for use by R
double PruneTerms[], // out: specifies which cols in bx are in best set
double RssVec[], // out: nTerms x 1
const int* pnCases, // in
const int* pnResp, // in: number of cols in y
const int* pnMaxTerms, // in
const double bx[], // in: MARS basis matrix, all cols must be indep
const double y[], // in: nCases * nResp (possibly weighted)
const double* pTrace) // in
{
TraceGlobal = *pTrace;
const int nMaxTerms = *pnMaxTerms;
bool* BoolPruneTerms = (bool*)malloc1(nMaxTerms * nMaxTerms * sizeof(bool),
"BoolPruneTerms\tMaxTerms %d nMaxTerms %d sizeof(bool) %d",
nMaxTerms, nMaxTerms, sizeof(bool));
size_t nCases = *pnCases; // type convert
EvalSubsetsUsingXtx(BoolPruneTerms, RssVec, nCases, *pnResp,
nMaxTerms, bx, y);
// convert BoolPruneTerms to upper triangular matrix PruneTerms
for(int iModel = 0; iModel < nMaxTerms; iModel++) {
int iPrune = 0;
for(int iTerm = 0; iTerm < nMaxTerms; iTerm++)
if(BoolPruneTerms[iTerm + iModel * nMaxTerms])
PruneTerms[iModel + iPrune++ * nMaxTerms] = iTerm + 1;
}
free1(BoolPruneTerms);
}
#endif // USING_R
//-----------------------------------------------------------------------------
#if STANDALONE && WEIGHTS
static void UnweightBx(
double bx[], // in: nCases x nMaxTerms
const double WeightsArg[], // in
const size_t nCases, // in: number of rows in bx
const int nMaxTerms) // in: number of cols in bx
{
if(WeightsArg) {
for(int iTerm = 0; iTerm < nMaxTerms; iTerm++)
for(int i = 0; i < (const int)nCases; i++)
bx_(i, iTerm) /= sqrt(WeightsArg[i]);
}
}
#endif // STANDALONE
//-----------------------------------------------------------------------------
#if STANDALONE
static void BackwardPass(
double* pBestGcv, // out: GCV of the best model i.e. BestSet columns of bx
bool BestSet[], // out: nMaxTerms x 1, indices of best set of cols of bx
double Residuals[], // out: nCases x nResp
double Betas[], // out: nMaxTerms x nResp
double bx[], // in: nCases x nMaxTerms, will be unweighted if weights
const double y[], // in: nCases x nResp
const double* WeightsArg, // in; NULL or nCases
const size_t nCases, // in: number of rows in bx and elements in y
const int nResp, // in: number of cols in y
const int nMaxTerms, // in: number of cols in bx
const double Penalty) // in: GCV penalty per knot
{
double* RssVec = (double*)malloc1(nMaxTerms * sizeof(double),
"RssVec\t\tnMaxTerms %d sizeof(double) %d",
nMaxTerms, sizeof(double));
bool* PruneTerms = (bool*)malloc1(nMaxTerms * nMaxTerms * sizeof(bool),
"PruneTerms\t\tnMaxTerms %d nMaxTerms %d sizeof(bool) %d",
nMaxTerms, nMaxTerms, sizeof(bool));
EvalSubsetsUsingXtx(PruneTerms, RssVec, nCases, nResp,
nMaxTerms, bx, y);
// now we have the RSS for each model, so find the iModel which has the best GCV
tprintf(3, "Backward pass:\nSubsetSize GRSq RSq\n");
int iBestModel = -1;
double GcvNull = GetGcv(1, nCases, RssVec[0], Penalty);
double BestGcv = POS_INF;
for(int iModel = 0; iModel < nMaxTerms; iModel++) {
const double Gcv = GetGcv(iModel+1, nCases, RssVec[iModel], Penalty);
if(Gcv < BestGcv) {
iBestModel = iModel;
BestGcv = Gcv;
}
double GRSq = 1 - BestGcv/GcvNull;
// Prevent negative almost-zero issued by some compilers (earth version 4.4.5)
// This prints as -0.0000 below and messes up the test scripts.
if(GRSq > -1e-12 && GRSq < 0) // -1e-12 is fairly arb
GRSq = 0;
tprintf(3, "%10d %12.4f %12.4f\n", iModel+IOFFSET,
GRSq, 1 - RssVec[iModel]/RssVec[0]);
}
tprintf(3, "\nBackward pass complete: selected %d terms of %d, "
"GRSq %.3f RSq %.3f\n\n",
iBestModel+IOFFSET, nMaxTerms,
1 - BestGcv/GcvNull, 1 - RssVec[iBestModel]/RssVec[0]);
// set BestSet to the model which has the best GCV
ASSERT(iBestModel >= 0);
memcpy(BestSet, PruneTerms + iBestModel * nMaxTerms, nMaxTerms * sizeof(bool));
free1(PruneTerms);
free1(RssVec);
*pBestGcv = BestGcv;
#if WEIGHTS
UnweightBx(bx, WeightsArg, nCases, nMaxTerms);
// TODO should use weighted regression in RegressAndFix below
#endif
// get final model Betas, Residuals, Rss
RegressAndFix(Betas, Residuals, NULL, BestSet,
bx, y, nCases, nResp, nMaxTerms);
}
#endif // STANDALONE
//-----------------------------------------------------------------------------
#if STANDALONE
static int DiscardUnusedTerms(
double bx[], // io: nCases x nMaxTerms
int Dirs[], // io: nMaxTerms x nPreds
double Cuts[], // io: nMaxTerms x nPreds
bool WhichSet[], // io: tells us which terms to discard
int nDegree[], // io: degree of each term, degree of intercept is 0
const int nMaxTerms,
const int nPreds,
const size_t nCases)
{
int nUsed = 0, iTerm;
for(iTerm = 0; iTerm < nMaxTerms; iTerm++)
if(WhichSet[iTerm]) {
memcpy(bx + nUsed * nCases, bx + iTerm * nCases, nCases * sizeof(double));
for(int iPred = 0; iPred < nPreds; iPred++) {
Dirs_(nUsed, iPred) = Dirs_(iTerm, iPred);
Cuts_(nUsed, iPred) = Cuts_(iTerm, iPred);
}
nDegree[nUsed] = nDegree[iTerm];
nUsed++;
}
memset(WhichSet, 0, nMaxTerms * sizeof(bool));
for(iTerm = 0; iTerm < nUsed; iTerm++)
WhichSet[iTerm] = true;
return nUsed;
}
#endif // STANDALONE
//-----------------------------------------------------------------------------
#if STANDALONE
void Earth(
double* pBestGcv, // out: GCV of the best model i.e. BestSet columns of bx
int* pnTerms, // out: max term nbr in final model, after removing lin dep terms
int* piTermCond, // out: reason we terminated the foward pass
bool BestSet[], // out: nMaxTerms x 1, indices of best set of cols of bx
double bx[], // out: nCases x nMaxTerms
int Dirs[], // out: nMaxTerms x nPreds, -1,0,1,2 for iTerm, iPred
double Cuts[], // out: nMaxTerms x nPreds, cut for iTerm, iPred
double Residuals[], // out: nCases x nResp
double Betas[], // out: nMaxTerms x nResp
const double x[], // in: nCases x nPreds
const double y[], // in: nCases x nResp
const double WeightsArg[], // in: nCases x 1, can be NULL, not yet supported
const size_t nCases, // in: number of rows in x and elements in y
const int nResp, // in: number of cols in y
const int nPreds, // in: number of cols in x
const int nMaxDegree, // in: Friedman's mi
const int nMaxTerms, // in: includes the intercept term
const double Penalty, // in: GCV penalty per knot
const double Thresh, // in: forward step threshold
const int nMinSpan, // in: set to non zero to override internal calculation
const int nEndSpan, // in: set to non zero to override internal calculation
const bool Prune, // in: do backward pass
const int nFastK, // in: Fast MARS K
const double FastBeta, // in: Fast MARS ageing coef
const double NewVarPenalty, // in: penalty for adding a new variable
const int LinPreds[], // in: nPreds x 1, 1 if predictor must enter linearly
const double AdjustEndSpan, // in:
const bool AutoLinPreds, // in: assume predictor linear if knot is min predictor value
const bool UseBetaCache, // in: 1 to use the beta cache, for speed
const double Trace, // in: 0 none 1 overview 2 forward 3 pruning 4 more pruning
const char* sPredNames[]) // in: predictor names in trace printfs, can be NULL
{
#if _MSC_VER && _DEBUG
InitMallocTracking();
#endif
TraceGlobal = Trace;
nMinSpanGlobal = nMinSpan;
nEndSpanGlobal = nEndSpan;
AdjustEndSpanGlobal = AdjustEndSpan;
// nUses is the number of time each predictor is used in the model
nUses = (int*)malloc1(nPreds * sizeof(int),
"nUses\t\t\tnPreds %d sizeof(int) %d",
nPreds, sizeof(int));
// nDegree is degree of each term, degree of intercept is considered to be 0
nDegree = (int*)malloc1(nMaxTerms * sizeof(int),
"nDegree\t\tnMaxTerms %d sizeof(int) %d",
nMaxTerms, sizeof(int));
double* yw = NULL;
#if WEIGHTS
if(WeightsArg) {
error("weights are not yet supported in STANDALONE earth"); // TODO
yw = (double*)malloc1(nCases * nResp * sizeof(double),
"yw\t\t\tnCases %d nResp %d sizeof(double) %d",
(const int)nCases, nResp, sizeof(double));
for(int iResp = 0; iResp < nResp; iResp++)
for(int i = 0; i < (const int)nCases; i++) {
const int j = iResp * (const int)nCases + i;
yw[j] = sqrt(WeightsArg[i]) * y[j];
}
}
#else
ASSERT(WeightsArg == NULL); // weights are not currently supported
#endif
int nTerms = 0, iTermCond = 0;
ForwardPass(&nTerms, &iTermCond,
BestSet, bx, Dirs, Cuts, nDegree, nUses,
x, y, yw, WeightsArg,
nCases, nResp, nPreds, nMaxDegree, nMaxTerms,
Penalty, Thresh, nFastK, FastBeta, NewVarPenalty,
LinPreds, AdjustEndSpan, AutoLinPreds, UseBetaCache, sPredNames);
// ensure bx is full rank by updating BestSet, and get Residuals and Betas
RegressAndFix(Betas, Residuals, NULL, BestSet,
bx, yw? yw: y, nCases, nResp, nMaxTerms);
if(TraceGlobal >= 6)
PrintSummary(nMaxTerms, nTerms, nPreds, nResp,
BestSet, Dirs, Cuts, Betas, nDegree);
int nMaxTerms1 = DiscardUnusedTerms(bx, Dirs, Cuts, BestSet, nDegree,
nMaxTerms, nPreds, nCases);
if(Prune)
BackwardPass(pBestGcv, BestSet, Residuals, Betas,
bx, yw? yw: y, WeightsArg, nCases, nResp, nMaxTerms1, Penalty);
else if(WeightsArg) {
// TODO should use weighted regression in RegressAndFix
// UnweightBx(bx, WeightsArg, nCases, nMaxTerms);
RegressAndFix(Betas, Residuals, NULL, BestSet,
bx, y, nCases, nResp, nMaxTerms);
}
if(TraceGlobal >= 6)
PrintSummary(nMaxTerms, nMaxTerms1, nPreds, nResp,
BestSet, Dirs, Cuts, Betas, nDegree);
*pnTerms = nMaxTerms1;
*piTermCond = iTermCond;
if(yw)
free1(yw);
free1(nDegree);
free1(nUses);
}
#endif // STANDALONE
//-----------------------------------------------------------------------------
// Return the max number of knots in any term.
// Lin dep factors are considered as having one knot (at the min value of the predictor)
#if STANDALONE
static int GetMaxKnotsPerTerm(
const bool UsedCols[], // in
const int Dirs[], // in
const int nPreds, // in
const int nTerms, // in
const int nMaxTerms) // in
{
int nKnotsMax = 0;
for(int iTerm = 1; iTerm < nTerms; iTerm++)
if(UsedCols[iTerm]) {
int nKnots = 0; // number of knots in this term
for(int iPred = 0; iPred < nPreds; iPred++)
if(Dirs_(iTerm, iPred) != 0)
nKnots++;
if(nKnots > nKnotsMax)
nKnotsMax = nKnots;
}
return nKnotsMax;
}
#endif // STANDALONE
//-----------------------------------------------------------------------------
// print a string representing the earth expression, one term per line
// TODO spacing is not quite right and is overly complicated
#if STANDALONE
static void FormatOneResponse(
const bool UsedCols[],// in: nMaxTerms x 1, indices of best set of cols of bx
const int Dirs[], // in: nMaxTerms x nPreds, -1,0,1,2 for iTerm, iPred
const double Cuts[], // in: nMaxTerms x nPreds, cut for iTerm, iPred
const double Betas[], // in: nMaxTerms x nResp
const int nPreds,
const int iResp,
const int nTerms,
const int nMaxTerms,
const int nDigits, // number of significant digits to print
const double MinBeta) // terms with fabs(beta) less than this are not printed, 0 for all
{
int iBestTerm = 0;
int nKnotsMax = GetMaxKnotsPerTerm(UsedCols, Dirs, nPreds, nTerms, nMaxTerms);
int nKnots = 0;
char s[1000];
ASSERT(nDigits >= 0);
char sFormat[50]; sprintf(sFormat, "%%-%d.%dg", nDigits+6, nDigits);
char sFormat1[50]; sprintf(sFormat1, "%%%d.%dg", nDigits+6, nDigits);
int nPredWidth;
if(nPreds > 100)
nPredWidth = 3;
else if(nPreds > 10)
nPredWidth = 2;
else
nPredWidth = 1;
char sPredFormat[20]; sprintf(sPredFormat, "%%%dd", nPredWidth);
char sPad[500]; sprintf(sPad, "%*s", 28+nDigits+nPredWidth, " "); // comment pad
const int nUsedCols = nTerms; // nUsedCols is needed for the Betas_ macro
printf(sFormat, Betas_(0, iResp)); // intercept
while(nKnots++ < nKnotsMax)
printf("%s", sPad);
printf(" // 0\n");
for(int iTerm = 1; iTerm < nTerms; iTerm++)
if(UsedCols[iTerm]) {
iBestTerm++;
if(fabs(Betas_(iBestTerm, iResp)) >= MinBeta) {
printf("%+-9.3g", Betas_(iBestTerm, iResp));
nKnots = 0;
for(int iPred = 0; iPred < nPreds; iPred++) {
switch(Dirs_(iTerm, iPred)) {
case 0:
break;
case -1:
sprintf(s, " * max(0, %s - %*sx[%s])",
sFormat, nDigits+2, " ", sPredFormat);
printf(s, Cuts_(iTerm, iPred), iPred);
nKnots++;
break;
case 1:
sprintf(s, " * max(0, x[%s]%*s- %s)",
sPredFormat, nDigits+2, " ", sFormat1);
printf(s, iPred, Cuts_(iTerm, iPred));
nKnots++;
break;
case 2:
sprintf(s, " * x[%s]%*s ",
sPredFormat, nDigits+2, " ");
printf(s, iPred);
nKnots++;
break;
default:
ASSERT(false);
break;
}
}
while(nKnots++ < nKnotsMax)
printf("%s", sPad);
printf(" // %d\n", iBestTerm);
}
}
}
void FormatEarth(
const bool UsedCols[],// in: nMaxTerms x 1, indices of best set of cols of bx
const int Dirs[], // in: nMaxTerms x nPreds, -1,0,1,2 for iTerm, iPred
const double Cuts[], // in: nMaxTerms x nPreds, cut for iTerm, iPred
const double Betas[], // in: nMaxTerms x nResp
const int nPreds,
const int nResp, // in: number of cols in y
const int nTerms,
const int nMaxTerms,
const int nDigits, // number of significant digits to print
const double MinBeta) // terms with fabs(betas) less than this are not printed, 0 for all
{
for(int iResp = 0; iResp < nResp; iResp++) {
if(nResp > 1)
printf("Response %d:\n", iResp+IOFFSET);
FormatOneResponse(UsedCols, Dirs, Cuts, Betas, nPreds, iResp,
nTerms, nMaxTerms, nDigits, MinBeta);
}
}
#endif // STANDALONE
//-----------------------------------------------------------------------------
// return the value predicted by an earth model, given a vector of inputs x
#if STANDALONE
static double PredictOneResponse(
const double x[], // in: vector nPreds x 1 of input values
const bool UsedCols[], // in: nMaxTerms x 1, indices of best set of cols of bx
const int Dirs[], // in: nMaxTerms x nPreds, -1,0,1,2 for iTerm, iPred
const double Cuts[], // in: nMaxTerms x nPreds, cut for iTerm, iPred
const double Betas[], // in: nMaxTerms x 1
const int nPreds, // in: number of cols in x
const int nTerms,
const int nMaxTerms)
{
double yHat = Betas[0];
int iTerm1 = 0;
for(int iTerm = 1; iTerm < nTerms; iTerm++)
if(UsedCols[iTerm]) {
iTerm1++;
double Term = Betas[iTerm1];
for(int iPred = 0; iPred < nPreds; iPred++) {
switch(Dirs_(iTerm, iPred)) {
case 0: break;
case -1: Term *= max(0, Cuts_(iTerm, iPred) - x[iPred]); break;
case 1: Term *= max(0, x[iPred] - Cuts_(iTerm, iPred)); break;
case 2: Term *= x[iPred]; break;
default: ASSERT("bad direction" == NULL); break;
}
}
yHat += Term;
}
return yHat;
}
void PredictEarth(
double y[], // out: vector nResp
const double x[], // in: vector nPreds x 1 of input values
const bool UsedCols[], // in: nMaxTerms x 1, indices of best set of cols of bx
const int Dirs[], // in: nMaxTerms x nPreds, -1,0,1,2 for iTerm, iPred
const double Cuts[], // in: nMaxTerms x nPreds, cut for iTerm, iPred
const double Betas[], // in: nMaxTerms x nResp
const int nPreds, // in: number of cols in x
const int nResp, // in: number of cols in y
const int nTerms,
const int nMaxTerms)
{
for(int iResp = 0; iResp < nResp; iResp++)
y[iResp] = PredictOneResponse(x, UsedCols, Dirs, Cuts,
Betas + iResp * nTerms, nPreds, nTerms, nMaxTerms);
}
#endif // STANDALONE
//-----------------------------------------------------------------------------
// Example main routine
// See earth/inst/slowtests/test.earthc.c for more complex examples
#if STANDALONE && MAIN
void error(const char *args, ...) // params like printf
{
char s[1000];
va_list va;
va_start(va, args);
vsprintf(s, args, va);
va_end(va);
printf("\nError: %s\n", s);
exit(-1);
}
// extern here prevents clang -Wmissing-prototypes warning
extern void xerbla_(char *srname, int* info);
void xerbla_(char *srname, int* info) // needed by BLAS and LAPACK routines
{
char buf[7];
strncpy(buf, srname, 6);
buf[6] = 0;
error("BLAS/LAPACK routine %6s gave error code %d", buf, -(*info));
}
int main(void)
{
const int nMaxTerms = 21; // called "nk" in the R code
const size_t nCases = 100; // note that nCases is size_t, not int
// this allows e.g. mallocs below to be bigger than 2GB
const int nResp = 1; // number of responses i.e. number of y columns
const int nPreds = 1; // number of predictors i.e. number of x columns
const int nMaxDegree = 1; // called "degree" in the R code
const double Penalty = (nMaxDegree > 1)? 3: 2;
const double Thresh = .001;
const int nMinSpan = 0; // 0 means auto
const int nEndSpan = 0; // 0 means auto
const bool Prune = true;
const int nFastK = 20;
const double FastBeta = 1;
const double NewVarPenalty = 0;
int* LinPreds = (int*)calloc1(nPreds, sizeof(int), NULL); // "linpreds" in R code
const double AdjustEndSpan = 2.0;
const bool AutoLinPreds = true;
const bool UseBetaCache = true;
const double Trace = 3;
const char** sPredNames = NULL;
double BestGcv;
int nTerms;
int iTermCond;
bool* BestSet = (bool*) malloc1(nMaxTerms * sizeof(bool), NULL);
double* bx = (double*)malloc1(nCases * nMaxTerms * sizeof(double), NULL);
int* Dirs = (int*) malloc1(nMaxTerms * nPreds * sizeof(int), NULL);
double* Cuts = (double*)malloc1(nMaxTerms * nPreds * sizeof(double), NULL);
double* Residuals = (double*)malloc1(nCases * nResp * sizeof(double), NULL);
double* Betas = (double*)malloc1(nMaxTerms * nResp * sizeof(double), NULL);
double* x = (double*)malloc1(nCases * nPreds * sizeof(double), NULL);
double* y = (double*)malloc1(nCases * nResp * sizeof(double), NULL);
ASSERT(nResp == 1); // code below only works for nResp == 1
for(int i = 0; i < (const int)nCases; i++) {
const double xi = i / (double)nCases;
x[i] = xi;
y[i] = sin(4 * xi); // target function, change this to whatever you want
}
Earth(&BestGcv, &nTerms, &iTermCond, BestSet, bx, Dirs, Cuts, Residuals, Betas,
x, y, NULL /*WeightsArg*/, nCases, nResp, nPreds,
nMaxDegree, nMaxTerms, Penalty, Thresh, nMinSpan, nEndSpan, Prune,
nFastK, FastBeta, NewVarPenalty, LinPreds,
AdjustEndSpan, AutoLinPreds, UseBetaCache, Trace, sPredNames);
printf("Expression:\n");
FormatEarth(BestSet, Dirs, Cuts, Betas, nPreds, nResp, nTerms, nMaxTerms, 3, 0);
double x1 = 0.1234, y1;
PredictEarth(&y1,
&x1, BestSet, Dirs, Cuts, Betas, nPreds, nResp, nTerms, nMaxTerms);
printf("\nf(%g) = %g\n", x1, y1);
free1(y);
free1(x);
free1(Betas);
free1(Residuals);
free1(Cuts);
free1(Dirs);
free1(bx);
free1(BestSet);
free1(LinPreds);
return 0;
}
#endif // STANDALONE && MAIN
|