/
StateSpaceContinuous.R
executable file
·491 lines (358 loc) · 15.7 KB
/
StateSpaceContinuous.R
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
#
# Copyright 2007-2019 by the individuals mentioned in the source code history
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#------------------------------------------------------------------------------
# Author: Michael D. Hunter
# Date: 2015-04-08
# Filename: StateSpaceContinuous.R
# Purpose: Create and test three example of continuous time state space models.
# Example 1 is an undamped linear oscillator.
# Example 2 is a damped linear oscillator with dynamic noise.
# Example 3 is a general dynamic model with multiple indicators per factor.
# All examples are fit to simulated data. Their parameter estimates are
# compared to their true generating values. This last example was also fit
# with a completely R-based implementation of the same algorithm that is now
# implemented in the C++ backend. The C++ code reaches the same solution
# as the R code, and is about 70x faster on the same machine.
#------------------------------------------------------------------------------
#------------------------------------------------------------------------------
# Load packages
require(mvtnorm)
require(Matrix)
require(OpenMx)
#------------------------------------------------------------------------------
# Example 1
# Undamped linear oscillator, i.e. a noisy sine wave.
# Measurement error, but no dynamic error, single indicator.
# This example works great.
#--------------------------------------
# Data Generation
set.seed(405)
tlen <- 200
t <- seq(1.2, 50, length.out=tlen)
freqParam <- .5
initialCond <- matrix(c(2.5, 0))
x <- initialCond[1,1]*cos(freqParam*t)
plot(t, x, type='l')
measVar <- 1.5
y <- cbind(obs=x+rnorm(tlen, sd=sqrt(measVar)), tim=t)
plot(t, y[,1], type='l')
#--------------------------------------
# Model Specification
#Note: the bounds are here only to keep CSOLNP from
# stepping too far off a cliff. With the bounds in
# place, CSOLNP finds the right solution. Without
# the bounds, CSOLNP goes crazy.
cdim <- list('obs', c('ksi', 'ksiDot'))
amat <- mxMatrix('Full', 2, 2, c(F, T, F, T), c(0, -.1, 1, -.2), name='A', lbound=-10)
bmat <- mxMatrix('Zero', 2, 1, name='B')
cmat <- mxMatrix('Full', 1, 2, FALSE, c(1, 0), name='C', dimnames=cdim)
dmat <- mxMatrix('Zero', 1, 1, name='D')
qmat <- mxMatrix('Zero', 2, 2, name='Q')
rmat <- mxMatrix('Diag', 1, 1, TRUE, .4, name='R', lbound=1e-6)
xmat <- mxMatrix('Full', 2, 1, TRUE, c(0, 0), name='x0', lbound=-10, ubound=10)
pmat <- mxMatrix('Diag', 2, 2, FALSE, 1, name='P0')
umat <- mxMatrix('Zero', 1, 1, name='u')
tmat <- mxMatrix('Full', 1, 1, name='time', labels='data.tim')
osc <- mxModel("LinearOscillator",
amat, bmat, cmat, dmat, qmat, rmat, xmat, pmat, umat, tmat,
mxExpectationStateSpace('A', 'B', 'C', 'D', 'Q', 'R', 'x0', 'P0', 'u', t='time', scores=TRUE),
mxFitFunctionML(),
mxData(y, 'raw'))
#osc <- mxOption(osc, 'Calculate Hessian', 'No')
#osc <- mxOption(osc, 'Standard Errors', 'No')
#osc <- mxOption(osc, 'Major iterations', 0)
oscr <- mxRun(osc)
#--------------------------------------
# Results Examination
summary(oscr)
(ssFreqParam <- mxEval(sqrt(-A[2,1]), oscr))
freqParam
omxCheckWithinPercentError(ssFreqParam, freqParam, 1)
(ssMeasVar <- mxEval(R, oscr))
measVar
omxCheckWithinPercentError(ssMeasVar, measVar, 10)
dampingParam <- 0
(ssDampingParam <- mxEval(-A[2,2], oscr))
dampingParam
omxCheckCloseEnough(ssDampingParam, dampingParam, .01)
rms <- function(x, y){sqrt(mean((x-y)^2))}
(ssInitialCond <- mxEval(x0, oscr))
initialCond
omxCheckTrue(rms(ssInitialCond, initialCond) < .1)
#------------------------------------------------------------------------------
# Example 1 continued
# Data generation and re-fitting
new_data <- mxGenerateData(oscr, nrow(y))
new_data$tim <- y[, 'tim']
nosc <- mxModel(osc, mxData(new_data, 'raw'))
noscr <- mxRun(nosc)
# Parameters look close
cbind(coef(oscr), coef(noscr))
se <- summary(oscr)$parameters[,6]
# Check that the old and new data estimates (data generated by sine+noise vs exact discrete)
# all with confidence intervals of one another.
withinCI <- (coef(noscr) < (coef(oscr) + 1.5*se)) & (coef(noscr) > (coef(oscr) - 1.5*se))
omxCheckEquals(withinCI, rep(TRUE, length(se)))
#------------------------------------------------------------------------------
# Compare frontend and backend scores
fstart <- Sys.time()
ks <- mxKalmanScores(noscr)
fstop <- Sys.time()
bstart <- Sys.time()
ksb <- mxKalmanScores(noscr, frontend=FALSE)
bstop <- Sys.time()
# Compare time to compute
fstop-fstart
bstop-bstart # 70x faster in backend
# Compare values
omxCheckTrue(all.equal(ks, ksb[1:8]))
# values equal to roughly 1.5e-8 #.Machine$double.eps ^ 0.5
cor(cbind(ks$xPredicted, noscr$expectation$xPredicted))
cor(cbind(ks$xUpdated, noscr$expectation$xUpdated))
cor(cbind(ks$xSmoothed, noscr$expectation$xSmoothed))
omxCheckCloseEnough( rms(ks$xPredicted, noscr$expectation$xPredicted), 0, 1e-10)
omxCheckCloseEnough( rms(ks$xUpdated, noscr$expectation$xUpdated), 0, 1e-10)
omxCheckCloseEnough( rms(ks$xSmoothed, noscr$expectation$xSmoothed), 0, 1e-10)
omxCheckCloseEnough( rms(t(apply(ks$PPredicted, 3, vech)), noscr$expectation$PPredicted), 0, 1e-10)
omxCheckCloseEnough( rms(t(apply(ks$PUpdated, 3, vech)), noscr$expectation$PUpdated), 0, 1e-10)
omxCheckCloseEnough( rms(t(apply(ks$PSmoothed, 3, vech)), noscr$expectation$PSmoothed), 0, 1e-10)
plot(new_data$tim, new_data$obs, type='l')
lines(new_data$tim, ks$xSmoothed[-1, 1], col='blue')
#------------------------------------------------------------------------------
# Example 2
# Damped linear oscillator example
# There is measurement noise and there are unmeasured dynamic disturbances.
# These disturbances are called dynamic noise.
# There is a single indicator.
# TODO Figure out what combination of variables can really be estimated for
# this kind of model. It is clear that measurement noise and SOME dynamic
# noise can be estimated, but not all.
#--------------------------------------
# Data Generation
xdim <- 2
udim <- 1
ydim <- 1
tdim <- 200
set.seed(315)
tA <- matrix(c(0, -.3, 1, -.7), xdim, xdim)
tB <- matrix(c(0), xdim, udim)
tC <- matrix(c(1, 0), ydim, xdim)
tD <- matrix(c(0), ydim, udim)
tQ <- matrix(c(0), xdim, xdim); diag(tQ) <- c(0, 2.2)
tR <- matrix(c(0), ydim, ydim); diag(tR) <- c(1.5)
x0 <- matrix(c(0, 1), xdim, 1)
P0 <- diag(c(1), xdim)
tdx <- matrix(0, xdim, tdim+1)
tx <- matrix(0, xdim, tdim+1)
tu <- matrix(0, udim, tdim)
ty <- matrix(0, ydim, tdim)
tT <- matrix((0:tdim)/4, nrow=1, ncol=tdim+1)
tI <- diag(1, nrow=xdim)
tx[,1] <- x0
for(i in 2:(tdim+1)){
q <- t(rmvnorm(1, rep(0, xdim), tQ))
tdx[,i] <- tA %*% tx[,i-1] + tB %*% tu[,i-1] + q
expA <- as.matrix(expm(tA * (tT[,i]-tT[,i-1])))
intA <- solve(tA) %*% (expA - tI)
eF <- as.matrix( expm( rbind(cbind(-tA, tQ), cbind(matrix(0, xdim, xdim), t(tA))) * (tT[,i]-tT[,i-1])))
tQd <- expA %*% eF[1:xdim, (xdim+1):(2*xdim)]
# Whiten the continuous-time error then transform it to the continuous time error cov
s <- svd(tQ)
dinv <- s$d
dinv[s$d > 0] <- 1/s$d[s$d > 0]
W <- diag(sqrt(dinv)) %*% t(s$v)
qd <- t(chol(tQd)) %*% W %*% q
tx[,i] <- expA %*% tx[, i-1] + intA %*% tB %*% tu[,i-1] + qd
#ty[,i-1] <- tC %*% tx[,i-1] + tD %*% tu[,i-1] + t(rmvnorm(1, rep(0, ydim), tR))
ty[,i-1] <- tC %*% tx[,i] + tD %*% tu[,i-1] + t(rmvnorm(1, rep(0, ydim), tR))
}
#plot(tx[1,], type='l')
rownames(ty) <- paste('y', 1:ydim, sep='')
rownames(tx) <- paste('x', 1:xdim, sep='')
plot(tT[,-1], ty[1,], type='l')
plot(tT[1,], tx[1,], type='l')
#--------------------------------------
# Model Specification
cdim <- list(rownames(ty), c('ksi', 'ksiDot'))
amat <- mxMatrix('Full', xdim, xdim, c(F, T, F, T), c(0, -.1, 1, -.2), name='A')
bmat <- mxMatrix('Zero', xdim, udim, name='B')
cmat <- mxMatrix('Full', ydim, xdim, FALSE, c(1, 0), name='C', dimnames=cdim)
dmat <- mxMatrix('Zero', udim, udim, name='D')
qmat <- mxMatrix('Diag', xdim, xdim, c(F,T), c(0, 1), name='Q', lbound=1e-6)
rmat <- mxMatrix('Diag', ydim, ydim, TRUE, 1.5, name='R')
xmat <- mxMatrix('Full', xdim, 1, c(T,F), c(0, 1), name='x0')
pmat <- mxMatrix('Diag', xdim, xdim, FALSE, 1, name='P0')
umat <- mxMatrix('Zero', udim, 1, name='u')
tmat <- mxMatrix('Full', 1, 1, name='time', labels='data.times')
dosc <- mxModel("DampedLinearOscillator",
amat, bmat, cmat, dmat, qmat, rmat, xmat, pmat, umat, tmat,
mxExpectationStateSpaceContinuousTime('A', 'B', 'C', 'D', 'Q', 'R', 'x0', 'P0', 'u', t='time'),
mxFitFunctionML(),
mxData(cbind(t(ty), times=tT[,-1]), 'raw'))
#dosc <- mxOption(dosc, 'Calculate Hessian', 'No')
#dosc <- mxOption(dosc, 'Standard Errors', 'No')
#dosc <- mxOption(dosc, 'Major iterations', 0)
doscr <- mxRun(dosc)
#--------------------------------------
# Results Examination
summary(doscr)
mxEval(A, doscr)
tA
omxCheckTrue(rms(mxEval(A, doscr)[2,1], tA[2,1]) < .07)
mxEval(R, doscr)
tR
omxCheckTrue(rms(mxEval(R, doscr), tR) < .2)
mxEval(Q, doscr)
tQ
omxCheckTrue(rms(mxEval(Q, doscr)[2,2], tQ[2,2]) < .7)
mxEval(x0, doscr) #poorly estimated, likely due to dynamic error
x0
omxCheckTrue(rms(mxEval(x0, doscr)[1,1], x0[1,1]) < .7)
estParam <- coef(doscr)
truParam <- c(tA[2,], tQ[2,2], tR, x0[1,1])
estSE <- summary(doscr)$parameters[,6]
ex3Tvals <- (estParam - truParam) / estSE
omxCheckTrue(all(abs(ex3Tvals) < 0.6))
s <- solve(kronecker(diag(xdim), mxEval(A, doscr)) + kronecker(mxEval(A, doscr), diag(xdim)))
(asym_lvar <- matrix(-s %*% matrix(mxEval(Q, doscr)), xdim, xdim))
(obs_lvar <- var(t(tx)))
omxCheckTrue(rms(vech(asym_lvar), vech(obs_lvar)) < .7)
(asym_ovar <- mxEval(C, doscr) %*% asym_lvar %*% t(mxEval(C, doscr)) + mxEval(R, doscr))
(obs_ovar <- var(t(ty)))
omxCheckTrue(rms(vech(asym_ovar), vech(obs_ovar)) < 1.1)
#------------------------------------------------------------------------------
# Example 3
# Full bore example of general dynamics with dynamic error,
# measurement error, and multiple indicators per latent variable.
#--------------------------------------
# Data generation
xdim <- 3
udim <- 2
ydim <- 9
tdim <- 500
set.seed(948)
tA <- matrix(c(-.4, 0, 0, 0, -.9, .1, 0, -.1, -.9), xdim, xdim)
tB <- matrix(c(0), xdim, udim)
tC <- matrix(c(runif(3, .4, 1), rep(0, ydim), runif(3, .4, 1), rep(0, ydim), runif(3, .4, 1)), ydim, xdim)
tD <- matrix(c(0), ydim, udim)
tQ <- matrix(c(0), xdim, xdim); diag(tQ) <- runif(xdim, 1, 2) # 0, 1
tR <- matrix(c(0), ydim, ydim); diag(tR) <- runif(ydim)
x0 <- matrix(c(rnorm(xdim)), xdim, 1)
P0 <- diag(c(runif(xdim)))
tdx <- matrix(0, xdim, tdim+1)
tx <- matrix(0, xdim, tdim+1)
tu <- matrix(0, udim, tdim)
ty <- matrix(0, ydim, tdim)
tT <- matrix(0:tdim, nrow=1, ncol=tdim+1)
tI <- diag(1, nrow=xdim)
tx[,1] <- x0
for(i in 2:(tdim+1)){
q <- t(rmvnorm(1, rep(0, xdim), tQ))
tdx[,i] <- tA %*% tx[,i-1] + tB %*% tu[,i-1] + q
expA <- as.matrix(expm(tA * (tT[,i]-tT[,i-1])))
intA <- solve(tA) %*% (expA - tI)
eF <- as.matrix( expm( rbind(cbind(-tA, tQ), cbind(matrix(0, xdim, xdim), t(tA))) * (tT[,i]-tT[,i-1])))
tQd <- expA %*% eF[1:xdim, (xdim+1):(2*xdim)]
# Whiten the continuous-time error then transform it to the continuous time error cov
s <- svd(tQ)
dinv <- s$d
dinv[s$d > 0] <- 1/s$d[s$d > 0]
W <- diag(sqrt(dinv)) %*% t(s$v)
qd <- t(chol(tQd)) %*% W %*% q
tx[,i] <- expA %*% tx[, i-1] + intA %*% tB %*% tu[,i-1] + qd
#ty[,i-1] <- tC %*% tx[,i-1] + tD %*% tu[,i-1] + t(rmvnorm(1, rep(0, ydim), tR))
ty[,i-1] <- tC %*% tx[,i] + tD %*% tu[,i-1] + t(rmvnorm(1, rep(0, ydim), tR))
}
#plot(tx[1,], type='l')
rownames(ty) <- paste('y', 1:ydim, sep='')
rownames(tx) <- paste('x', 1:xdim, sep='')
plot(ty[1,], type='l')
#--------------------------------------
# Model Specification
smod <- mxModel(
name='SSContinuous',
mxMatrix(name='A', values=tA, nrow=xdim, ncol=xdim, free=c(TRUE, FALSE, FALSE, FALSE, TRUE, TRUE, FALSE, FALSE, TRUE), labels=c('a', NA, NA, NA, 'b', 'c', NA, 'csym[1,1]', 'b'), ubound=c(0, NA, NA, NA, 0, NA, NA, NA, 0)),
mxAlgebra(name='csym', -c),
mxMatrix(name='B', values=0, nrow=xdim, ncol=udim, free=FALSE),
mxMatrix(name='C', values=tC, nrow=ydim, ncol=xdim, free=(tC!=0), dimnames=list(rownames(ty), rownames(tx))),
mxMatrix(name='D', values=0, nrow=ydim, ncol=udim, free=FALSE),
# Note Factor error matrix is fixed! This is for model identification.
# I happen to fix the variances to their true values.
mxMatrix(name='Q', type='Diag', values=diag(tQ), nrow=xdim, ncol=xdim, free=FALSE),
mxMatrix(name='R', type='Diag', values=diag(tR), nrow=ydim, ncol=ydim, free=TRUE),
mxMatrix(name='x', values=x0, nrow=xdim, ncol=1, free=FALSE),
mxMatrix(name='P', values=P0, nrow=xdim, ncol=xdim, free=FALSE),
mxMatrix("Zero", udim, 1, name="u"),
mxMatrix("Full", 1, 1, name="time", labels="data.Time"),
mxData(observed=cbind(t(ty), Time=tT[,-1]), type='raw'),
mxExpectationSSCT(A='A', B='B', C='C', D='D', Q='Q', R='R', x0='x', P0='P', u='u', t='time'),
mxFitFunctionML()
)
# Uncomment for degugging
#smod <- mxOption(smod, 'Calculate Hessian', 'No')
#smod <- mxOption(smod, 'Standard Errors', 'No')
#smod <- mxOption(smod, 'Major iterations', 0)
srun <- mxRun(smod)
# takes about a minute on Squishy.
# The same model all in R takes 1.46 hours.
# Speed-up is approx 70x.
# Check likelihoods of initial parameters
# -2LL
#summary(srun)$Minus2LogLikelihood # when major iterations is 0
#2*dlmLL(y=t(ty), mod=mfun(tinit)) + 200*9*log(2*pi) # dlm gives back -LL - CONST, so adjust it. 200 is N, 9 is k
#--------------------------------------
# Results Examination
summary(srun)
mxEval(A, srun); tA
omxCheckTrue(rms(mxEval(A, srun)[tA!=0], tA[tA!=0]) < .02)
mxEval(C, srun)[tC!=0]; tC[tC!=0]
omxCheckTrue(rms(mxEval(C, srun)[tC!=0], tC[tC!=0]) < .1)
diag(mxEval(R, srun)); diag(tR)
omxCheckTrue(rms(diag(mxEval(R, srun)), diag(tR)) < .1)
restA <- matrix(c(
-0.3173239, 0.00000000, 0.00000000,
0.0000000, -1.09531962, -0.04313248,
0.0000000, 0.04313248, -1.09531962
), xdim, xdim, byrow=TRUE)
restC <- c(0.4789963, 0.6038891, 0.7602543, 0.9514004, 0.7272259,
1.0570274, 0.5431997, 1.0255853, 0.4885580)
restR <- c(0.3226227, 0.8767227, 1.1029966, 0.5607970, 0.1622717,
0.8304199, 0.7078884, 0.9453716, 0.9422717)
restAnew <- matrix(c(
-0.4381936, 0, 0,
0, -0.90029646, -0.08843642,
0, 0.08843642, -0.90029646), xdim, xdim, byrow=TRUE)
restCnew <- c(0.4779426, 0.7274788, 0.7311741, 0.9677682, 0.7325752,
1.0499232, 0.4789363, 1.0284996, 0.5794153)
restRnew <- c(0.3115990, 0.8528872, 1.1730305, 0.5813367, 0.1538722,
0.8300052, 0.7317202, 0.8460866, 0.9378043)
# At least 7x improvement in dynamics estimation
omxCheckTrue(rms(restA, tA) / rms(restAnew, tA) > 7)
# At most 30% decrement in factor loadings estimation
omxCheckTrue(rms(restCnew, tC[tC!=0]) / rms(restC, tC[tC!=0]) < 1.3)
# At most 20% decrement in residual variance estimation
omxCheckTrue(rms(restRnew, diag(tR)) / rms(restR, diag(tR)) < 1.2)
omxCheckCloseEnough(diag(mxEval(A, srun)), diag(restAnew), 0.05)
omxCheckCloseEnough(mxEval(C, srun)[tC!=0], restCnew, 0.01)
omxCheckCloseEnough(diag(mxEval(R, srun)), restRnew, 0.02)
s <- solve(kronecker(diag(xdim), mxEval(A, srun)) + kronecker(mxEval(A, srun), diag(xdim)))
(asym_lvar <- matrix(-s %*% matrix(mxEval(Q, srun)), xdim, xdim))
(obs_lvar <- var(t(tx)))
omxCheckTrue(rms(vech(asym_lvar), vech(obs_lvar)) < .05)
(asym_ovar <- mxEval(C, srun) %*% asym_lvar %*% t(mxEval(C, srun)) + mxEval(R, srun))
(obs_ovar <- var(t(ty)))
omxCheckTrue(rms(vech(asym_ovar), vech(obs_ovar)) < 0.05)
#------------------------------------------------------------------------------
# Done