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
|
<html>
<head>
<title>Solvers</title>
</head>
<body>
<h1>Solvers</h1>
<p>
In this section we enumerate and briefly describe each of Cain's solvers.
See the <a href="Simulation.htm">Simulation Methods</a> section for more details
and derivations. In the <a href="UserMethod.htm">method editor</a> the
solvers are grouped according to four heirarchical categories: time
dependence, simulation output, algorithm, and option. One choose a
solver by selecting items from each of these in turn.
</p>
<a name="TimeDependence"><h2>Time Dependence</h2></a>
<p>
The first category differentiates between time homogeneous and time
inhomogeneous systems. In a time homogenous problem the reaction
propensities do not have explicit time dependence. Time inhomogenous
problems typically arise when the volume or temperature have a
prescribed time dependence. This implies that the reaction rates will
explicitly depend on time. Cain has many solvers for time homogeneous
problems, but few for time inhomogenous ones. While conceptually very
similar, one uses different algorithms for the two cases. Thus in
choosing a solver one first selects the time dependence category.
</p>
<a name="SimulationOutput"><h2>Simulation Output</h2></a>
<p>
Next one selects the desired simulation output. The various kinds of output
are described in the <a href="UserMethod.htm">Method Editor</a> section.
To recap, one can either collect time series or histogram data.
For time series data one can either record the state at uniformly spaced
time frames or one can record every reaction event. Although Cain is
primarily for stochastic simulations, it does offer a few simple deterministic
solvers. These generate time series data and are grouped in a separate
category. When the species populations are recorded in histograms,
the mean and standard deviation are also recorded. This is useful because
unless the bin width is unity one can only approximately calculate these
statistics from the histogram data.
For histogram output one can either study transient or steady
state behavior. For the former one records the state at uniformly spaced
frames in time. For the latter one averages the populations over
a specified time interval. Thus, in summary there are five simulation output
categories:
<ul>
<li>
<tt>Time Series, Uniform</tt>.
Here one takes snapshots of the state of uniformly spaced frames in time.
One can use either exact methods like the
<a href="SimulationDirect.htm">direct method</a>,
<a href="SimulationFirst.htm">first reaction method</a>, and
<a href="SimulationNext.htm">next reaction method</a>,
or approximate methods such as
<a href="SimulationTau.htm">tau-leaping</a> and
<a href="SimulationHybrid.htm">hybrid direct/tau-leaping</a>.
Recently there has been much research in optimized exact methods.
(See reference [<a href="Bibliography.htm#mauch2011">Mauch 2011</a>]
for an analysis.) Thus we
offer many different formulations of the direct and next reaction methods.
You may experiment to see how the various formulations perform. However,
the default formulation has good performance for most problems.
<li>
<tt>Time Series, All Reactions</tt>.
Recording every reaction event is useful when one wants a visual
representation of the amount of noise in a system. By contrast, when one
records the state at frames one does not get a picture of the magnitude of
the random fluctuations. Note that recording every reaction is expensive,
and plotting the results is even more so. For this output option we use
the 2-D search formulation of the direct method.
<li>
<tt>Time Series, Deterministic</tt>.
Deterministic systems are modeled with a set of ordinary differential
equations. Cain uses <a href="SimulationOde.htm">numerical integration</a>
to approximately solve the system. The state is recorded at uniformly
spaced time frames.
<li>
<tt>Histograms, Transient Behavior</tt>.
If one records the state at time frames one can of course plot the
trajectories. One can also analyze the species populations at these
frames. One can compute the mean and standard deviation of the population
or one can plot the distribution of populations in a histogram. Obtaining
accurate statistics or distribution data usually requires generating many
trajectories. In this case, the state information may require a lot of
storage capacity. It is more economical to dynamically update population
statistics and a histograms of the species populations as the trajectories
are being generated. For this output category populations statistics and
histogram data is stored at uniformly spaced frames in time. To generate
the trajectories we use the 2-D search formulation of the direct method.
<li>
<tt>Histograms, Steady State</tt>.
If one records every reaction event, one can analyze the steady state
behavior of a system. One could take the time average of the species
populations to determine the mean and standard deviation of the steady
state populations. One could also record the time weighted populations
in a histogram to determine the distribution of steady state populations.
However, storing every reaction event typically requires a lot of storage.
It is more efficient to dynamically update statistics and histogram data
as the trajectories are generated. For this output option the time averaged
statistics and histogram data is collected. Again we use the 2-D search
formulation of the direct method to generate the trajectories.
</ul>
</p>
<!--CONTINUE Show little plots for each output option.-->
<a name="Algorithms"><h2>Algorithms</h2></a>
<p>
The third category in selecting a solvers is the simulation algorithm.
There are algorithms for exact and approximate stochastic simulation.
Below is a list with some implementation details.
</p>
<ul>
<li>
<tt>Direct</tt>.
Gillespie's
<a href="SimulationDirect.htm">direct method</a>
[<a href="Bibliography.htm#gillespie1976">Gillespie 1976</a>]
[<a href="Bibliography.htm#gillespie1977">Gillespie 1977</a>].
A reaction dependency graph
[<a href="Bibliography.htm#gibson2000">Gibson 2000</a>]
is used to minimize the cost of recomputing propensities.
Unless otherwise indicated, the sum of the propensities and associated
quantities are dynamically updated.
At each step the time increment is determined by drawing an exponential
deviate. The mean of the distribution is the sum of the propensities.
The reaction is chosen by
generating a discrete deviate whose weighted probability mass function
is the set of reaction propensities. There are many options for
generating the discrete deviate. The 2-D search formulation is the default
method.
<li>
<tt>Next Reaction</tt>.
Gibson and Bruck's
<a href="SimulationNext.htm">next reaction method</a>
[<a href="Bibliography.htm#gibson2000">Gibson 2000</a>] stores reaction
times in a priority queue. There are several options for implementing
this data structure.
<li>
<tt>First Reaction</tt>.
Gillespie's
<a href="SimulationFirst.htm">first reaction method</a>
[<a href="Bibliography.htm#gillespie1976">Gillespie 1976</a>]
[<a href="Bibliography.htm#gillespie1977">Gillespie 1977</a>].
<li>
<tt>Tau-Leaping</tt>.
We implement explicit <a href="SimulationTau.htm">tau-leaping</a>
with the step size selection presented
in reference [<a href="Bibliography.htm#cao2006a">Cao 2006a</a>].
<li>
<tt>Hybrid Direct/Tau-Leaping</tt>.
This hybrid algorithm uses the direct method for the volatile (low
population) species and tau-leaping for the rest.
<li>
<tt>ODE, Integrate Reactions</tt>.
For <a href="SimulationOde.htm">ODE integration</a>
we use
<a href="http://en.wikipedia.org/wiki/Runge-Kutta">Runge-Kutta</a>
schemes.
</ul>
<a name="Options"><h2>Options</h2></a>
<p>
Below is a list of the available solvers grouped according to the
heirarchical categories.
Each of the solvers use sparse arrays for the state change vectors
[<a href="Bibliography.htm#li2006">Li 2006</a>].
The methods that require exponential deviates use the ziggurat method
[<a href="Bibliography.htm#marsaglia2000">Marsaglia 2000</a>].
<ul>
<li>
<h3><tt>Time Homogeneous</tt></h3>
<ul>
<li>
<h4><tt>Time Series, Uniform</tt></h4>
<ul>
<li>
<h5><tt>Direct</tt></h5>
The computational complexities for generating discrete deviates
and for modifying propensities are indicated for each method in
terms of the number of reactions <em>M</em>.
<ul>
<li>
<tt>2-D search</tt>
- Generate O(<em>M<sup>1/2</sup></em>). Modify O(1).
The propensities are stored in a 2-D table that stores the sum of each
row. The number of rows (and hence the number of columns) is
O(<em>M<sup>1/2</sup></em>). The discrete deviate is generated by
first determining the appropriate row with a linear search and then
performing a linear search within that row
[<a href="Bibliography.htm#mauch2011">Mauch 2011</a>].
<li>
<tt>2-D search, sorted</tt>
- Generate O(<em>M<sup>1/2</sup></em>). Modify O(1).
The propensities are periodically ordered so the larger values have
smaller
<a href="http://en.wikipedia.org/wiki/Manhattan_distance">Manhattan
distance</a> from the lower corner of the table than smaller
values
[<a href="Bibliography.htm#mauch2011">Mauch 2011</a>].
For some problems this may reduce the costs of
searching.
<li>
<tt>2-D search, bubble sort</tt>
- Generate O(<em>M<sup>1/2</sup></em>). Modify O(1).
The propensities are dynamically ordered each time they are modified
[<a href="Bibliography.htm#mauch2011">Mauch 2011</a>].
<li>
<tt>Composition rejection</tt>
- Generate O(1). Modify O(1).
Propensities are placed in groups. In each
group the propensities differ by at most a factor of two. A deviate
is generated by a linear search on the groups followed by selecting
a reaction within a group with the method of rejection.
This solver implements a slightly modified version of the composition
rejection method presented in [<a href="Bibliography.htm#slepoy2008">Slepoy 2008</a>].
<li>
<tt>Binary search, full CMF</tt>
- Generate O(log <em>M</em>). Modify O(<em>M</em>).
The cumulative mass function (CMF) is stored in an array. This allows
one to generate a discrete deviate with a binary search on that array.
At each time step the CMF is regenerated. This is an implementation
of the logarithmic direct method [<a href="Bibliography.htm#li2006">Li 2006</a>].
<li>
<tt>Binary search, sorted CMF</tt>
- Generate O(log <em>M</em>). Modify O(<em>M</em>).
In this solver the reactions are arranged in descending order
according to the sum of the propensities of the influencing
reactions
[<a href="Bibliography.htm#mauch2011">Mauch 2011</a>].
This minimizes the portion of the CMF that one has
to recompute after firing a reaction and updating the
influenced propensities.
<li>
<tt>Binary search, recursive CMF</tt>
- Generate O(log <em>M</em>). Modify O(log <em>M</em>).
Instead of storing the full CMF, a partial, recursive CMF is used.
This approach is equivalent to the method presented in
[<a href="Bibliography.htm#gibson2000">Gibson 2000</a>]. A deviate can be generated
with a binary search. Modifying a propensity affects at most
log<sub>2</sub> <em>M</em> elements of the partial, recursive CMF.
<li>
<tt>Linear search</tt>
- Generate O(<em>M</em>). Modify O(1).
We use a linear search on the PMF to generate a deviate. The sum of
the propensities is dynamically updated.
<li>
<tt>Linear search, delayed update</tt>
- Generate O(<em>M</em>). Modify O(1).
This solver recomputes the sum of the propensities at each time step
as done in the original formulation of the direct method
[<a href="Bibliography.htm#gillespie1977">Gillespie 1977</a>].
<li>
<tt>Linear search, sorted</tt>
- Generate O(<em>M</em>). Modify O(1).
The propensities are periodically sorted in descending order
[<a href="Bibliography.htm#mccollum2005">McCollum 2005</a>].
<li>
<tt>Linear search, bubble sort</tt>
- Generate O(<em>M</em>). Modify O(1).
The propensities are dynamically ordered by using swaps each time
a propensity is modified
[<a href="Bibliography.htm#mauch2011">Mauch 2011</a>].
</ul>
<li>
<h5><tt>Next Reaction</tt></h5>
Each formulation below uses a different data structure to
implement the priority queue. The computational complexities for
removing an element and modifying an element are given in terms
of the number of reactions <em>M</em>.
<ul>
<li>
<tt>Hashing</tt>
- Pop O(1). Modify O(1).
The indexed priority queue is implemented with a hash table
[<a href="Bibliography.htm#cormen2001">Cormen 2001</a>,
<a href="Bibliography.htm#mauch2011">Mauch 2011</a>]. We use hashing
with chaining,
which means that each bin in the hash table contains a list of
elements. The hash function is a linear function of the putative
reaction times, truncated to an integer to obtain a bin index.
The hash function is dynamically adapted to obtain the target load.
(The <em>load</em> is the average number of elements in each bin.)
<li>
<tt>Binary heap, pointer</tt>
- Pop O(log <em>M</em>). Modify O(log <em>M</em>).
This solver uses a binary heap that is implemented with an array
of pointers to the putative reaction times. This is the data structure
used in [<a href="Bibliography.htm#gibson2000">Gibson 2000</a>].
<li>
<tt>Binary heap, pair</tt>
- Pop O(log <em>M</em>). Modify O(log <em>M</em>).
We use a binary heap implemented with an array of index/time pairs
[<a href="Bibliography.htm#mauch2011">Mauch 2011</a>].
The pair version of the binary heap typically has better performance
than the pointer version on 64-bit architectures. Vice versa for
32-bit architectures.
<li>
<tt>Partition</tt>
- Pop O(<em>M<sup>1/2</sup></em>). Modify O(<em>M<sup>1/2</sup></em>).
A splitting value is used to divide the reactions two categories
[<a href="Bibliography.htm#mauch2011">Mauch 2011</a>].
The splitting value is chosen so that initially there are
O(<em>M<sup>1/2</sup></em>) reactions in the lower partition.
The remainder are in the upper partition. In determining the minimum
putative reaction time, one only has to examine reactions in the
lower partition. When the lower partition is empty, a new splitting
value is chosen.
<li>
<tt>Linear search</tt>
- Pop O(<em>M</em>). Modify O(1).
The putative reaction times are stored in an array. We use a linear
search to find the minimum
[<a href="Bibliography.htm#mauch2011">Mauch 2011</a>].
</ul>
<li>
<h5><tt>First Reaction</tt></h5>
Gillespie's first reaction method
[<a href="Bibliography.htm#gillespie1976">Gillespie 1976</a>]
[<a href="Bibliography.htm#gillespie1977">Gillespie 1977</a>].
<ul>
<li>
<tt>Simple</tt>
- A simple implementation of the method.
<li>
<tt>Reaction influence</tt>
- We use the reaction influence data structure to minimize recomputing
the propensities.
<li>
<tt>Absolute time</tt>
- We use absolute times instead of waiting times for each reaction.
This allows us to reduce the number of exponential deviates used.
After firing a reaction, we only generate new reaction times for the
influenced reactions.
</ul>
<li>
<h5><tt>Tau-Leaping</tt></h5>
All of the tau-leaping solvers have first order accuracy. They differ in
how they calculate the expected propensities, whether they correct
negative populations, and how they calculate the time step. For calculating
the expected propensities one may use a first, second, or fourth order
Runge-Kutta scheme. (The first and second order schemes are commonly
called forward and midpoint, respectively.)
While all yield a first order stochastic method, using
a higher order scheme is typically more efficient. Use the adaptive step
size solvers unless you are studying the effect of step size. With a fixed
step size it is difficult to predict the accuracy of the simulation.
With the adaptive step size solvers you may choose whether or not to
correct negative populations. Without correction, some portion of the
simulations may fail.
<ul>
<li>
<tt>Runge-Kutta, 4th order</tt>
<li>
<tt>Midpoint</tt>
<li>
<tt>Forward</tt>
<li>
<tt>Runge-Kutta, 4th order, no correction</tt>
<li>
<tt>Midpoint, no correction</tt>
<li>
<tt>Forward, no correction</tt>
<li>
<tt>Runge-Kutta, 4th order, fixed step size</tt>
<li>
<tt>Midpoint, fixed step size</tt>
<li>
<tt>Forward, fixed step size</tt>
</ul>
<li>
<h5><tt>Hybrid Direct/Tau-Leaping</tt></h5>
With these solvers you can choose the scheme for calculating the expected
propensities.
<ul>
<li>
<tt>Runge-Kutta, 4th order</tt>
<li>
<tt>Midpoint</tt>
<li>
<tt>Forward</tt>
</ul>
</ul>
<li>
<h4><tt>Time Series, All Reactions</tt></h4>
The direct method with a 2-D search is used when recording each reaction
event.
<ul>
<li>
<h5><tt>Direct</tt></h5>
<ul>
<li>
<tt>2-D search</tt>
</ul>
</ul>
<li>
<h4><tt>Time Series, Deterministic</tt></h4>
When generating a single deterministic trajectory one can either use a
built-in solver or export the problem to a Mathematica notebook.
The solvers use
<a href="http://en.wikipedia.org/wiki/Runge-Kutta">Runge-Kutta</a>
schemes. For ordinary work, use the
adaptive step size
<a href="http://en.wikipedia.org/wiki/Cash-Karp">Cash-Karp</a> variant,
which is a fifth order method. The rest of the solvers are only useful
in studying how the order and step size affect the solution.
<ul>
<li>
<h5><tt>ODE, Integrate Reactions</tt></h5>
<ul>
<li>
<tt>Runge-Kutta, Cash-Karp</tt>
<li>
<tt>Runge-Kutta, Cash-Karp, fixed step size</tt>
<li>
<tt>Runge-Kutta, 4th order, fixed step size</tt>
<li>
<tt>Midpoint, fixed step size</tt>
<li>
<tt>Forward, fixed step size</tt>
</ul>
<li>
<h5><tt>Mathematica</tt></h5>
<ul>
<li>
<tt>NDSolve</tt>
</ul>
</ul>
<li>
<h4><tt>Histograms, Transient Behavior</tt></h4>
<ul>
<li>
<h5><tt>Direct</tt></h5>
<ul>
<li>
<tt>2-D search</tt>
</ul>
</ul>
<li>
<h4><tt>Histograms, Steady State</tt></h4>
By recording the average value of species populations you can determine
the steady state solution (if it exists).
<ul>
<li>
<h5><tt>Direct</tt></h5>
Each of the following use the direct method with a 2-D search.
<ul>
<li>
<tt>Elapsed time</tt> -
This is the most efficient method for most problems. The algorithm
keeps track of when species change values in order to minimize the cost
of recording the populations.
<li>
<tt>Time steps</tt> -
This is the traditional method. The species populations are recorded
at every time step.
<li>
<tt>All possible states</tt> -
This is an implementation of the all possible states method
[<a href="Bibliography.htm#lipshtat2007">Lipshtat 2007</a>].
It is usually less efficient than the traditional method.
</ul>
</ul>
</ul>
<li>
<h3><tt>Time Inhomogeneous</tt></h3>
<ul>
<li>
<h4><tt>Time Series, Uniform</tt></h4>
<ul>
<li>
<h5><tt>Direct</tt></h5>
<ul>
<li>
<h6><tt>Constant propensities</tt></h6>
</ul>
</ul>
<li>
<h4><tt>Time Series, All Reactions</tt></h4>
<ul>
<li>
<h5><tt>Direct</tt></h5>
<ul>
<li>
<h6><tt>Constant propensities</tt></h6>
</ul>
</ul>
<li>
<h4><tt>Time Series, Deterministic</tt></h4>
<ul>
<li>
<h5><tt>ODE, Integrate Reactions</tt></h5>
<ul>
<li>
<h6><tt>Runge-Kutta, Cash-Karp</tt></h6>
</ul>
</ul>
<li>
<h4><tt>Histograms, Transient Behavior</tt></h4>
<ul>
<li>
<h5><tt>Direct</tt></h5>
<ul>
<li>
<h6><tt>Constant propensities</tt></h6>
</ul>
</ul>
<li>
<h4><tt>Histograms, Stead State</tt></h4>
<ul>
<li>
<h5><tt>Direct</tt></h5>
<ul>
<li>
<h6><tt>Constant propensities</tt></h6>
</ul>
</ul>
</ul>
</ul>
</p>
</body>
</html>
|