-
Notifications
You must be signed in to change notification settings - Fork 2
/
process.R
426 lines (392 loc) · 20.6 KB
/
process.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
##############################################
### ###
### Adaptive Design of Experiments for ###
### Maximizing Information Gain ###
### code by: Brennan Klein ###
### ###
##############################################
## Initializing the landscape and search process.
randomRows <- function(datagrid,n_init=25,sobol="Sobol") {
## During initialization, this function is called to generate
## the initial samples of points on the information surface.
if (sobol=="Sobol") {
return(getSobol(datagrid, n_init))
}
samp <- datagrid[sample(nrow(datagrid), n_init),]
if (any(duplicated(samp)==T)) {
samp <- datagrid[sample(nrow(datagrid), n_init),]
}
return(samp)
}
getSobol <- function(datagrid, n_init) {
## Generates a random Sobol Sequence.
## this is a quasi-random sequence, with a random seed.
df <- rbind(c(0,0), sobol(n_init, 2, scrambling=3, seed=sample(c(1:1000), 1)))
colnames(df) <-c('PI', 'A')
df <- data.frame(df)
sampy <- min(datagrid$A) + df$A * (max(datagrid$A) - min(datagrid$A))
sampx <- min(datagrid$PI) + df$PI* (max(datagrid$PI)- min(datagrid$PI))
out <- data.frame("PI"=sampx,"A"=sampy)
out <- out[-1,]
rownames(out) <- 1:nrow(out)
return(out)
}
initialize_process <- function(posteriors, n_init, expParam1_range,
expParam2_range, n_dim, sampling_method,
n_players, n_rounds, n_samples, sobol, sampling,
asymmetric, L, models_used, current_top_model=1) {
## First: Call this function to get the initial dataframe and the
## corresponding information numbers for each point in the surface.
datagrid <- expand.grid(A=seq(expParam1_range[1], expParam1_range[2],
length.out=n_dim),
PI=seq(expParam2_range[1],
expParam2_range[2], length.out=n_dim))
history <- randomRows(datagrid, n_init, sobol)
history$I <- 0
history$Stop <- 2
history$Round <- c(1:dim(history)[1])
str <- paste0("Initializing %02i %s seeds with %05s sampled histories per ",
"search ... Current Round is: %02i ... Current Time is: %s")
LIMIT <- nrow(history) - L + 1
for (row in 1:LIMIT) {
a <- history[row,]$A
p <- history[row,]$PI
history[row,]$I <- getInfoNumber(a, p, sampling_method, posteriors,
n_samples, asymmetric, sampling,
models_used, n_players, n_rounds,
current_top_model)
print(sprintf(str, n_init, sobol, n_samples, row, Sys.time()))
}
GPredict_prev <- run_gp(history[1:LIMIT,], datagrid)
for (row in (LIMIT+1):nrow(history)) {
a <- history[row,]$A
p <- history[row,]$PI
history[row,]$I <- getInfoNumber(a, p, sampling_method, posteriors,
n_samples, asymmetric, sampling,
models_used, n_players, n_rounds,
current_top_model)
GPredict_curr <- run_gp(history[1:row,], datagrid)
history[row,]$Stop <- stopping_criteria(GPredict_prev, GPredict_curr)
GPredict_prev <- GPredict_curr
print(sprintf(str, n_init, sobol, n_samples, row, Sys.time()))
}
return(history)
}
###################################
### Gaussian Process functions. ###
###################################
## Gaussian process
run_algorithm_once <- function(history, datagrid, posteriors, rho0, failsafe,
L, n_samples, k, asymmetric, sampling, search,
n_init, sampling_method, sobol, seed,
models_used, current_top_model=1) {
## Iterates through the Gaussian Process functions,
## searching more and more spaces.
## Info string.
istr <- "Current Round is: %03i (%s) ... Stop is: %04f ... Currently stopping at: %s"
current_round <- dim(history)[1]
if (search == "GPUCBPE") {
history <- add_gpucb_pe_point(history, datagrid, current_round,
n_samples, posteriors, k, L, asymmetric,
sampling, rho0, sampling_method,
models_used, current_top_model)
current_round <- dim(history)[1]
while (current_round < failsafe) {
history <- add_gpucb_pe_point(history, datagrid, current_round,
n_samples, posteriors, k, L,
asymmetric, sampling, rho0,
sampling_method, models_used,
current_top_model)
current_round <- dim(history)[1]
stop <- all(tail(history,L)$Stop > (1-rho0))
if (stop) {
print(sprintf(istr, current_round, search,
round(tail(history,1)$Stop, 5), Sys.time()))
history$sampling_method <- sampling_method
history$Search <- search
history$n_init <- n_init
history$n_samples <- n_samples
history$Sobol <- sobol
history$Sampling <- sampling
history$csvname <- make_filename(asymmetric, sampling, n_init,
sampling_method, n_samples,
search, seed, sobol,
models_used, current_top_model,
dtype="CSV")
history$pngname <- make_filename(asymmetric, sampling, n_init,
sampling_method, n_samples,
search, seed, sobol,
models_used, current_top_model,
dtype="FIG")
gp_seq <- c("UCB",rep("PE", k))
times <- trunc(dim(history)[1]/length(gp_seq))
plus <- dim(history)[1] %% length(gp_seq)
history$SearchLabel <- c(rep(gp_seq, times),head(gp_seq,plus))
return(history)
}
}
print(sprintf(istr, current_round, search,
round(tail(history,1)$Stop, 5), Sys.time()))
history$sampling_method <- sampling_method
history$Search <- search
history$n_init <- n_init
history$n_samples <- n_samples
history$Sobol <- sobol
history$Sampling <- sampling
history$csvname <- make_filename(asymmetric, sampling, n_init,
sampling_method, n_samples, search,
seed, sobol, models_used,
current_top_model, dtype="CSV")
history$pngname <- make_filename(asymmetric, sampling, n_init,
sampling_method, n_samples, search,
seed, sobol, models_used,
current_top_model, dtype="FIG")
gp_seq <- c("UCB",rep("PE", k))
times <- trunc(dim(history)[1]/length(gp_seq))
plus <- dim(history)[1] %% length(gp_seq)
history$SearchLabel <- c(rep(gp_seq, times), head(gp_seq,plus))
return(history)
}
## Random.
current_round <- dim(history)[1]
history <- add_random_point(history, datagrid, current_round, n_samples,
posteriors, k, L, asymmetric, sampling, rho0,
search, n_init, sampling_method, models_used,
current_top_model)
current_round <- dim(history)[1]
while(current_round < failsafe) {
history <- add_random_point(history, datagrid, current_round,n_samples,
posteriors, k, L, asymmetric, sampling,
rho0, search, n_init, sampling_method,
models_used, current_top_model)
current_round <- dim(history)[1]
stop <- all(tail(history,L)$Stop > (1-rho0))
if (stop) {
print(sprintf(istr, current_round, search, round(tail(history,1)$Stop, 5), Sys.time()))
history$sampling_method <- sampling_method
history$Search <- search
history$n_init <- n_init
history$n_samples <- n_samples; history$Sobol <- sobol; history$Sampling <- sampling
history$csvname <- make_filename(asymmetric, sampling, n_init,
sampling_method, n_samples,
search, seed, sobol, models_used,
current_top_model, dtype="CSV")
history$pngname <- make_filename(asymmetric, sampling, n_init,
sampling_method, n_samples, search,
seed, sobol, models_used,
current_top_model, dtype="FIG")
history$SearchLabel <- "Rand"
return(history)
}
}
print(sprintf(istr, current_round, serch, round(tail(history,1)$Stop, 5), Sys.time()))
history$sampling_method <- sampling_method; history$Search <- search; history$n_init <- n_init
history$n_samples <- n_samples; history$Sobol <- sobol; history$Sampling <- sampling
history$csvname <- make_filename(asymmetric, sampling, n_init,
sampling_method, n_samples, search, seed,
sobol, models_used, current_top_model,
dtype="CSV")
history$pngname <- make_filename(asymmetric, sampling, n_init,
sampling_method, n_samples, search, seed,
sobol, models_used, current_top_model,
dtype="FIG")
history$SearchLabel <- "Rand"
return(history)
}
run_algorithm_grid <- function(datagrid, posteriors, n_samples, asymmetric,
sampling, sampling_method, seed,
models_used=c(1,2,3,4), current_top_model=1) {
istr <- "Current Round is: %03i (Grid) ... Just dug at: (%04f, %04f) ... Current Time is: %s"
k <- 0
n_init <- 0
search <- "Grid"
sobol <- "Grid"
history <- datagrid
history$I <- 0
history$Round <- c(1:dim(datagrid)[1])
history$Stop <- 2
for (row in history$Round) {
A <- history[row,]$A
PI <- history[row,]$PI
history[row,]$I <- getInfoNumber(A, PI, sampling_method, posteriors, n_samples, asymmetric, sampling, models_used, n_players, n_rounds, current_top_model)
print(sprintf(istr, row, round(history[row,]$PI, 5), round(history[row,]$A, 5), Sys.time()))
}
history$sampling_method <- sampling_method; history$Search <- search; history$n_init <- n_init
history$n_samples <- n_samples; history$Sobol <- sobol; history$Sampling <- sampling
history$csvname <- make_filename(asymmetric, sampling, n_init, sampling_method, n_samples, search, seed, sobol, models_used, current_top_model, dtype="CSV")
history$pngname <- make_filename(asymmetric, sampling, n_init, sampling_method, n_samples, search, seed, sobol, models_used, current_top_model, dtype="FIG")
history$SearchLabel <- "Grid"
return(history)
}
add_random_point <- function(history, datagrid, current_round, n_samples, posteriors, k, L, asymmetric, sampling, rho0, search, n_init, sampling_method, current_top_model=1) {
## Adds one new random point to history.
istr <- "Current Round is: %03i (Random) ... Just dug at: (%04f, %04f) ... Stop is: %04f ... Current Time is: %s"
GPredict_prev <- run_gp(history, datagrid)
new_point <- data.frame(PI=runif (1,.2,.8), A=runif (1,2,6))
new_point$I <- getInfoNumber(new_point$A, new_point$PI, sampling_method, posteriors, n_samples, asymmetric, sampling, models_used, n_players, n_rounds, current_top_model)
new_point$Round <- current_round+1; new_point$Stop <- 2;
history_curr <- rbind(history, new_point)
GPredict_curr <- run_gp(history_curr, datagrid)
current_round <- dim(history_curr)[1]
history_curr$Stop[current_round] <- stopping_criteria(GPredict_prev, GPredict_curr)
history <- history_curr
GPredict_prev <- GPredict_curr
print(sprintf(istr,
current_round, round(history[current_round-1,]$PI, 5), round(history[current_round-1,]$A, 5), round(tail(history,1)$Stop, 5), Sys.time()))
return(history)
}
add_gpucb_pe_point <- function(history, datagrid, current_round, n_samples, posteriors, k, L, asymmetric, sampling, rho0, sampling_method, models_used, current_top_model=1) {
## Does two things: First, a point is added that is the max(y_hat)+UCB. Then it adds k PE points.
## UCB step
istr <- "Current Round is: %03i (UCB) ... Just dug at: (%04f, %04f) ... Stop is: %04f ... Current Time is: %s"
GPredict_prev <- run_gp(history, datagrid)
ucb <- add_ucb_point(GPredict_prev, history, datagrid, current_round, n_samples, posteriors, asymmetric, sampling, sampling_method, models_used, current_top_model)
history <- ucb[[1]]
GPredict_prev <- ucb[[2]]
current_round <- ucb[[3]]
print(sprintf(istr,
current_round, round(history[current_round,]$PI, 5), round(history[current_round,]$A, 5), round(tail(history,1)$Stop, 5), Sys.time()))
stop <- all(tail(history,L)$Stop > (1-rho0))
if (stop) {
return(history)
}
## PE step
for (i in 1:k) {
pe <- add_pe_point(GPredict_prev, history, datagrid, current_round, n_samples, posteriors, asymmetric, sampling, sampling_method, models_used, current_top_model)
history <- pe[[1]]
GPredict_prev <- pe[[2]]
current_round <- pe[[3]]
kind <- pe[[4]]
print(sprintf("Current Round is: %03i %s... Just dug at: (%04f, %04f) ... Stop is: %04f ... Current Time is: %s",
current_round, kind, round(history[current_round,]$PI, 5), round(history[current_round,]$A, 5), round(tail(history,1)$Stop, 5), Sys.time()))
stop <- all(tail(history,L)$Stop > (1-rho0))
if (stop) {
return(history)
}
}
return(history)
}
jitter_points <- function(coords, datagrid, history) {
jitterA <- (unique(datagrid$A)[2] - unique(datagrid$A )[1])/2
jitterPI <- (unique(datagrid$PI)[2] - unique(datagrid$PI)[1])/2
if (typeof(coords)!="list") {
newA <- coords[1] + runif (1, -jitterA, jitterA)
newPI <- coords[2] + runif (1, -jitterPI, jitterPI)
coords <- c(newA, newPI)
} else {
newA <- coords$A + runif (1, -jitterA, jitterA)
newPI <- coords$PI + runif (1, -jitterPI, jitterPI)
for (round in c(1:dim(history)[1])) {
if (newPI[round] > max(datagrid$PI)) {
newPI[round] <- max(datagrid$PI)
} else if (newPI[round] < min(datagrid$PI)) {
newPI[round] <- min(datagrid$PI)
}
if (newA[round] > max(datagrid$A)) {
newA[round] <- max(datagrid$A)
} else if (newA[round] < min(datagrid$A)) {
newA[round] <- min(datagrid$A)
}
}
coords <- data.frame("A"=newA, "PI"=newPI)
}
return(coords)
}
add_ucb_point <- function(GPredict_prev, history, datagrid, current_round, n_samples, posteriors, asymmetric, sampling, sampling_method, models_used, current_top_model=1) {
## Adds one new UCB point to history.
y_plus_sigma <- GPredict_prev$Y_hat + 2*(GPredict_prev$MSE)**.5
UCB_point <- GPredict_prev$complete_data[which.max(y_plus_sigma),]
coords <- as.numeric(UCB_point[1:2])
new_point <- data.frame(PI=coords[2], A=coords[1])
new_point$I <- getInfoNumber(new_point$A, new_point$PI, sampling_method, posteriors, n_samples, asymmetric, sampling, models_used, n_players, n_rounds, current_top_model)
new_point$Round <- current_round+1; new_point$Stop <- 2
history_curr <- rbind(history, new_point)
GPredict_curr <- run_gp(history_curr, datagrid)
current_round <- dim(history_curr)[1]
history_curr$Stop[current_round] <- stopping_criteria(GPredict_prev, GPredict_curr)
return(list(history_curr, GPredict_curr, current_round))
}
add_pe_point <- function(GPredict_prev, history, datagrid, current_round, n_samples, posteriors, asymmetric, sampling, sampling_method, models_used, current_top_model=1) {
## Adds one new PE point to history.
y_minus_sigma <- GPredict_prev$Y_hat - 2*(GPredict_prev$MSE)**.5
y_plus_sigma <- GPredict_prev$Y_hat + 2*(GPredict_prev$MSE)**.5
region <- which(y_plus_sigma > max(y_minus_sigma))
if (length(region) > 1) {
space_to_sample <- GPredict_prev$complete_data[region,]
PE_point <- space_to_sample[which.max(space_to_sample[,4]),]
coords <- as.numeric(PE_point[1:2])
new_point <- data.frame(PI=coords[2], A=coords[1])
new_point$I <- getInfoNumber(new_point$A, new_point$PI, sampling_method, posteriors, n_samples, asymmetric, sampling, models_used, n_players, n_rounds, current_top_model)
new_point$Round <- current_round+1; new_point$Stop <- 2
history_curr <- rbind(history, new_point)
GPredict_curr <- run_gp(history_curr, datagrid)
current_round <- dim(history_curr)[1]
history_curr$Stop[current_round] <- stopping_criteria(GPredict_prev, GPredict_curr)
kind <- "(PE) "
} else {
PE_point <- GPredict_prev$complete_data[which.max(GPredict_prev$MSE),]
coords <- as.numeric(PE_point[1:2])
new_point <- data.frame(PI=coords[2], A=coords[1])
new_point$I <- getInfoNumber(new_point$A, new_point$PI, sampling_method, posteriors, n_samples, asymmetric, sampling, models_used, n_players, n_rounds, current_top_model)
new_point$Round <- current_round+1; new_point$Stop <- 2
history_curr <- rbind(history, new_point)
GPredict_curr <- run_gp(history_curr, datagrid)
current_round <- dim(history_curr)[1]
history_curr$Stop[current_round] <- stopping_criteria(GPredict_prev, GPredict_curr)
kind <- "(PER) "
}
return(list(history_curr, GPredict_curr, current_round, kind))
}
run_gp <- function(history, datagrid) {
## Takes the current history and creates a landscape using predict and GP_fit.
temp <- history
## temp$A <- round(temp$A, 3); temp$PI <- round(temp$PI, 3)
temp$A <- round(temp$A, 6)
temp$PI <- round(temp$PI, 6)
history_curr <- suppressWarnings(aggregate(temp, by=list(temp$A, temp$PI), FUN=mean))
anorm <- (history_curr$A - min(datagrid$A)) / (max(datagrid$A) - min(datagrid$A))
pnorm <- (history_curr$PI - min(datagrid$PI)) / (max(datagrid$PI) - min(datagrid$PI))
evidence <- data.frame("A"= anorm, "PI"=pnorm, "I"= history_curr$I)
gp_model <- GP_fit(evidence[,c("A", "PI")], evidence$I)
gp_model$X <- as.matrix(history_curr[,c("A","PI")])
GPredict <- predict(gp_model, datagrid)
return(GPredict)
}
stopping_criteria <- function(GPredict_prev, GPredict_curr) {
## Stop when the procedure ceases to learn about the landscape, when comparing
## the global changes in mu between two successive iterations.
prev_df <- data.frame(GPredict_prev$complete_data)
curr_df <- data.frame(GPredict_curr$complete_data)
prev_df$num <- c(1:dim(prev_df)[1])
curr_df$num <- c(1:dim(curr_df)[1])
pi_prev_sorted <- prev_df[order(prev_df$Y_hat),]
pi_curr_sorted <- curr_df[order(curr_df$Y_hat),]
pi_prev_rank <- pi_prev_sorted$num
pi_curr_rank <- pi_curr_sorted$num
numerator <- discounted_rank_dissimilarity(pi_curr_rank, pi_prev_rank)
denominator <- get_max_distance(length(pi_curr_rank))
rhoXv <- 1 - numerator/denominator
return(rhoXv)
}
get_rank_distances <- function(pi_t1, pi_t0) {
## Returns the distance between two ordered lists.
return(c(pi_t1 - pi_t0)**2)
}
discounted_rank_dissimilarity <- function(pi_t1, pi_t0) {
## Calculates the rank dissimilarity.
numerator <- get_rank_distances(pi_t1, pi_t0)
denominator <- pi_t1**2
dists <- numerator/denominator
d <- sum(dists)
return(d)
}
get_max_distance <- function(nv) {
## Find the maximum distance.
curr_max <- nv
for (i in 1:nv) {
normal <- c(1:i); revers <- c(i:1)
d <- discounted_rank_dissimilarity(normal, revers)
if (d > curr_max) curr_max <- d
}
return(curr_max)
}