-
Notifications
You must be signed in to change notification settings - Fork 0
/
val2.jl
403 lines (270 loc) · 13 KB
/
val2.jl
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
using LinearAlgebra, Statistics, Plots, Distributions, StatsBase, DelimitedFiles, CSV, Tables, Random, BenchmarkTools
using BayesianOptimization, GaussianProcesses, Distributions
BLAS.set_num_threads(1)
using TimerOutputs
to = TimerOutput()
#const mts = MersenneTwister.(1:Threads.nthreads())
# Parameters
const context_dim = 2
const context_mean = 0
const context_sd = 1
const obs_sd = 1
const bandit_count = 3
const bandit_prior_mean = 0
const bandit_prior_sd = 10
const T = 100
const n_episodes = 2
const discount = 1.
const epsilon = .01
const rollout_length = 100
const n_rollouts = 50000
const n_opt_rollouts = 10
const n_spsa_iter = 10
const n_grid_iter = 7
const grid_ratio = 2
const grid_num = 6
const int_length = 2
const n_grid_rollouts = 50
const n_points = 3
## SIMULATOR FUNCTION
const idx = Base.parse(Int, ENV["SLURM_ARRAY_TASK_ID"])
function grid_contextual_bandit_simulator(n_points, action_function, T, rollout_length, n_episodes, n_rollouts, n_opt_rollouts, context_dim, context_mean,
context_sd, obs_sd, bandit_count, bandit_prior_mean, bandit_prior_sd, discount, epsilon)
means = zeros(n_points, bandit_count, context_dim)
covs = zeros(n_points, bandit_count, context_dim, context_dim)
GRID_REWARDS = zeros(n_episodes*n_points, T)
GRID_POST_MEANS = zeros(n_episodes*n_points, T, bandit_count, context_dim)
GRID_POST_COVS = zeros(n_episodes*n_points, T, bandit_count, context_dim, context_dim)
for i in 1:n_points
for j in 1:bandit_count
means[i, j, :] = rand(MvNormal(repeat([bandit_prior_mean], context_dim),
Matrix((bandit_prior_sd^2)I,context_dim,context_dim)))
covs[i, j, :, :] = rand(InverseWishart(context_dim + 2, Matrix((bandit_prior_sd^2)I, context_dim, context_dim)))
end
end
for i in 1:n_points
bandit_prior_means = means[i, :, :]
bandit_prior_covs = covs[i, :, :, :]
TOT_REWARDS, TOT_POST_MEANS, TOT_POST_COVS = contextual_bandit_simulator(action_function, T, rollout_length, n_episodes, n_rollouts, n_opt_rollouts, context_dim, context_mean,
context_sd, obs_sd, bandit_count, bandit_prior_means, bandit_prior_covs, discount, epsilon)
#print(TOT_REWARDS)
GRID_REWARDS[((i-1)*n_episodes+1):(i*n_episodes), :] = TOT_REWARDS
GRID_POST_MEANS[((i-1)*n_episodes+1):(i*n_episodes), :, :, :] = TOT_POST_MEANS
GRID_POST_COVS[((i-1)*n_episodes+1):(i*n_episodes), :, :, :, :] = TOT_POST_COVS
end
return GRID_REWARDS, GRID_POST_MEANS, GRID_POST_COVS
end
function ep_contextual_bandit_simulator(ep,action_function, T, rollout_length, n_episodes, n_rollouts, n_opt_rollouts, context_dim, context_mean,
context_sd, obs_sd, bandit_count, bandit_prior_means, bandit_prior_covs, discount, epsilon, global_bandit_param)
#for ep in 1:n_episodes
#pct_done = round(100 * ep_count / n_episodes, digits=1)
#print(pct_done)
#threadreps[Threads.threadid()] += 1
##print("\rEpisode: $pct_done%")
#bandit_param = randn(bandit_count, context_dim) * bandit_prior_sd .+ bandit_prior_mean
bandit_posterior_means = zeros(bandit_count, context_dim)
bandit_posterior_covs = zeros(bandit_count, context_dim, context_dim)
bandit_param = copy(global_bandit_param)
true_bandit_param = copy(global_bandit_param)
EPREWARDS = zeros(T)
EPOPTREWARDS = zeros(T)
POST_MEANS = zeros(T, bandit_count, context_dim)
POST_COVS = zeros(T, bandit_count, context_dim, context_dim)
copy!(bandit_posterior_means, bandit_prior_means)
copy!(bandit_posterior_covs, bandit_prior_covs)
#io = open("~/rl/monitor/monitor_$(idx).txt", "w")
#write(io, 0)
#close(io)
for t in 1:T
POST_MEANS[t, :, :] = bandit_posterior_means
POST_COVS[t, :, :, :] = bandit_posterior_covs
context = randn(context_dim) * context_sd .+ context_mean
true_expected_rewards = true_bandit_param * context
action = action_function(t, T, bandit_count, context, bandit_posterior_means, bandit_posterior_covs, discount, epsilon, rollout_length, n_rollouts, n_opt_rollouts, context_dim)
true_expected_reward = true_expected_rewards[action]
EPREWARDS[t] = true_expected_reward
EPOPTREWARDS[t] = maximum(true_expected_rewards)
obs = randn() * obs_sd + true_expected_reward
#old_cov = bandit_posterior_covs[action, :, :]
#print(bandit_posterior_covs[action,:,:])
#old_precision = inv(bandit_posterior_covs[action,:,:])
old_cov = bandit_posterior_covs[action, :, :]
CovCon = old_cov * context ./ obs_sd
#bandit_posterior_covs[action, :, :] = inv(context * context' / obs_sd^2 + old_precision)
bandit_posterior_covs[action, :, :] = old_cov - CovCon * CovCon' ./ (1 + dot(context, old_cov, context))
bandit_posterior_covs[action, :, :] = ((bandit_posterior_covs[action,:,:]) + (bandit_posterior_covs[action,:,:])')/2
bandit_posterior_means[action, :] = (bandit_posterior_covs[action, :, :]) * (old_cov \ (bandit_posterior_means[action,:]) + context * obs / obs_sd^2)
println("Ep: ", ep, " - ", t, " of ", T, " for ", String(Symbol(action_function)))
flush(stdout)
#io = open("~/rl/monitor/monitor_$(idx).txt", "r")
#curr_prog = read(io, Int)
#close(io)
#io = open("~/rl/monitor/monitor_$(idx).txt", "w")
#write(io, curr_prog+1)
#close(io)
end
return EPREWARDS, POST_MEANS, POST_COVS
end
function contextual_bandit_simulator(action_function, T, rollout_length, n_episodes, n_rollouts, n_opt_rollouts, context_dim, context_mean,
context_sd, obs_sd, bandit_count, bandit_prior_means, bandit_prior_covs, discount, epsilon)
REWARDS = zeros(T, n_episodes)
OPTREWARDS = zeros(T, n_episodes)
TOT_REWARDS = zeros(n_episodes, T)
TOT_POST_MEANS = zeros(n_episodes, T, bandit_count, context_dim)
TOT_POST_COVS = zeros(n_episodes, T, bandit_count, context_dim, context_dim)
ep_count = 1
global_bandit_param = zeros(bandit_count, context_dim)
#threadreps = zeros(Threads.nthreads())
for ep in 1:n_episodes
for bandit in 1:bandit_count
global_bandit_param[bandit, :] = rand(MvNormal((bandit_prior_means[bandit,:]), (bandit_prior_covs[bandit,:,:])))
end
EPREWARDS, POST_MEANS, POST_COVS = ep_contextual_bandit_simulator(ep,action_function, T, rollout_length, n_episodes, n_rollouts, n_opt_rollouts, context_dim, context_mean,context_sd, obs_sd, bandit_count, bandit_prior_means, bandit_prior_covs, discount, epsilon, global_bandit_param)
ep_count += 1
TOT_REWARDS[ep, :] = EPREWARDS
TOT_POST_MEANS[ep, :, :, :] = POST_MEANS
TOT_POST_COVS[ep, :, :, :, :] = POST_COVS
end
#print(threadreps)
return TOT_REWARDS, TOT_POST_MEANS, TOT_POST_COVS
end
## POLICY FUNCTIONS
# Greedy
function greedy_policy(t, T, bandit_count, context, bandit_posterior_means, bandit_posterior_covs, discount, epsilon, rollout_length, n_rollouts, n_opt_rollouts, context_dim)
val, action = findmax(bandit_posterior_means * context)
return action
end
# Epsilon Greedy
function epsilon_greedy_policy(t, T, bandit_count, context, bandit_posterior_means, bandit_posterior_covs, discount, epsilon, rollout_length, n_rollouts, n_opt_rollouts, context_dim)
epsilon_draw = rand()
if epsilon_draw < epsilon
return rand(1:bandit_count)
else
val, action = findmax(bandit_posterior_means * context)
return(action)
end
end
# Bayes UCB
function bayes_ucb_policy(t, T, bandit_count, context, bandit_posterior_means, bandit_posterior_covs, discount, epsilon, rollout_length, n_rollouts, n_opt_rollouts, context_dim)
val, action = findmax(bandit_posterior_means * context)
reward_means = bandit_posterior_means * context
reward_sds = sqrt.(vec([context' * (@view bandit_posterior_covs[i,:,:]) * context for i=1:bandit_count]))
ucbs = quantile.(Normal.(reward_means, reward_sds), 1-1/t)
return findmax(ucbs)[2]
end
# Thompson
function thompson_policy(t, T, bandit_count, context, bandit_posterior_means, bandit_posterior_covs, discount, epsilon, rollout_length, n_rollouts, n_opt_rollouts, context_dim)
thompson_samples = zeros(bandit_count, context_dim)
for bandit in 1:bandit_count
thompson_samples[bandit, :] = rand(MvNormal((bandit_posterior_means[bandit,:]), (bandit_posterior_covs[bandit,:,:])))
end
return findmax(thompson_samples * context)[2]
end
# SquareCB
function squarecb_policy(t, T, bandit_count, context, bandit_posterior_means, bandit_posterior_covs, discount, epsilon, rollout_length, n_rollouts, n_opt_rollouts, context_dim)
reward_means = bandit_posterior_means * context
opt_mean, opt_id = findmax(reward_means)
selection_probs = zeros(bandit_count)
for bandit in 1:bandit_count
selection_probs[bandit] = 1 / (bandit_count + sqrt(bandit_count*T) * (opt_mean - reward_means[bandit]))
end
selection_probs[opt_id] = 0
selection_probs[opt_id] = 1 - sum(selection_probs)
return sample(weights(selection_probs))
end
## SIMULATIONS
# Sims
print("\n")
@time GRID_REWARDS, GRID_POST_MEANS, GRID_POST_COVS = grid_contextual_bandit_simulator(n_points, greedy_policy, T, rollout_length, n_episodes, n_rollouts, n_opt_rollouts, context_dim, context_mean,
context_sd, obs_sd, bandit_count, bandit_prior_mean, bandit_prior_sd, discount, epsilon)
#print(GRID_REWARDS)
cov_vec_len = convert(Int64, (context_dim + 1) * context_dim / 2)
output = zeros(T, n_points*n_episodes, 1 + (context_dim + cov_vec_len) * bandit_count)
upper_triangular_vec = function(M)
d = size(M)[1]
output = zeros(convert(Int64,(d+1) * d / 2))
count = 1
for i in 1:d
for j in i:d
output[count] = M[i, j]
count += 1
end
end
return output
end
for t in 1:T
cum_rewards = [sum(GRID_REWARDS[j, (T-t+1):T]) for j in 1:(n_points*n_episodes)]
output[t, :, 1] = cum_rewards
for bandit in 1:bandit_count
output[t, :, (2+context_dim*(bandit-1)):(1+context_dim*bandit)] =
GRID_POST_MEANS[:, (T-t+1), bandit, :]
for j in 1:(n_points*n_episodes)
#output[t, j, (2+context_dim*bandit_count+context_dim*(bandit-1)):(1+context_dim*bandit_count+context_dim*bandit)] = eigvals(GRID_POST_COVS[j, (T-t+1), bandit, :, :])
output[t, j, (2+context_dim*bandit_count+cov_vec_len*(bandit-1)):(1+context_dim*bandit_count+cov_vec_len*bandit)] = upper_triangular_vec(GRID_POST_COVS[j, (T-t+1), bandit, :, :])
end
end
#CSV.write("/hpc/home/jml165/rl/valresults/results_$(idx)_$(t).csv", Tables.table(output[t, :, :]))
end
using Flux
using BSON: @save, @load
if idx == 1
rm("/hpc/home/jml165/rl/valneuralnets/", recursive = true, force = true)
mkdir("/hpc/home/jml165/rl/valneuralnets")
@timeit to "loop" begin
for t in 1:T
dat = output[t,:,:]
ncol = size(dat)[2]
X = dat[:, 3:ncol]'
Y = dat[:, 2]'
opt = Descent(.01)
hidden_size = div(ncol-2, 2)
m = Chain(Dense(ncol - 2, hidden_size, σ), Dense(hidden_size, 1))
loss(x,y) = sum(Flux.Losses.mse(m(x),y))
@timeit to "train_$(t)" Flux.train!(loss, Flux.params(m), [(X,Y)], opt)
@timeit to "save" @save "/hpc/home/jml165/rl/valneuralnets/valnn_$(t)_$(idx).bson" m
end
end
show(to)
else
t = 1
dat = output[t,:,:]
ncol = size(dat)[2]
X = dat[:, 3:ncol]'
Y = dat[:, 2]'
opt = Descent(.01)
hidden_size = div(ncol-2, 2)
#@timeit to "load" @load "/hpc/home/jml165/rl/valneuralnets/valnn_$(t)_$(idx-1).bson" m
m = Chain(Dense(ncol - 2, hidden_size, σ), Dense(hidden_size, 1))
loss(x,y) = sum(Flux.Losses.mse(m(x),y))
Flux.train!(loss, Flux.params(m), [(X,Y)], opt)
#@timeit to "save" @save "/hpc/home/jml165/rl/valneuralnets/valnn_$(t)_$(idx).bson" m
#@timeit to "rm" rm("/hpc/home/jml165/rl/valneuralnets/valnn_$(t)_$(idx-1).bson")
while true
if isfile("/hpc/home/jml165/rl/valneuralnets/valnn_$(T)_$(idx-1).bson")
@timeit to "loop" begin
for t in 1:T
dat = output[t,:,:]
ncol = size(dat)[2]
X = dat[:, 3:ncol]'
Y = dat[:, 2]'
opt = Descent(.01)
hidden_size = div(ncol-2, 2)
@timeit to "load" @load "/hpc/home/jml165/rl/valneuralnets/valnn_$(t)_$(idx-1).bson" m
#m = Chain(Dense(ncol - 2, hidden_size, σ), Dense(hidden_size, 1))
loss(x,y) = sum(Flux.Losses.mse(m(x),y))
@timeit to "train_$(t)" Flux.train!(loss, Flux.params(m), [(X,Y)], opt)
@timeit to "save" @save "/hpc/home/jml165/rl/valneuralnets/valnn_$(t)_$(idx).bson" m
@timeit to "rm" rm("/hpc/home/jml165/rl/valneuralnets/valnn_$(t)_$(idx-1).bson")
end
end
show(to)
break
end
end
end
if idx == 399
for t in 1:T
@save "/hpc/home/jml165/rl/valneuralnets/valnn_$(t).bson" m
rm("/hpc/home/jml165/rl/valneuralnets/valnn_$(t)_$(idx).bson")
end
end