-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathhillcar_mcts.py
343 lines (306 loc) · 12.4 KB
/
hillcar_mcts.py
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
import numpy as np
import scipy as sp
import scipy.sparse as sps
import multiprocessing
import matplotlib.pyplot as plt
from mdp import *
from config.mdp import *
from config.solver import *
from utils import *
from discrete import *
from lcp import *
from utils import *
from experiment import *
#WORKERS = 1
WORKERS = multiprocessing.cpu_count()-1
BATCHES_PER_WORKER = 4
STATES_PER_BATCH = 20
SIM_HORIZON = 1000
BUILD_MODE = 'load'
#BUILD_MODE = 'build'
low_dim = 16
ref_dim = 64
BUDGETS = [4,8,16,32,64,128,256,512]
#BUDGETS = [32]
ROOT = os.path.expanduser('~/data/hillcar') # root filename
DRIVER = os.path.expanduser('~/repo/lcp-research/cdiscrete/driver')
SAVE_FILE = ROOT + '/hillcar.'
def build_problem(disc_n):
# disc_n = number of cells per dimension
step_len = 0.01 # Step length
n_steps = 2 # Steps per iteration
damp = 1e-4 # Dampening
jitter = 0.05 # Control jitter
discount = 0.997 # Discount (\gamma)
bounds = [[-2,6],[-4,4]] # Square bounds,
cost_radius = 0.25 # Goal region radius
actions = np.array([[-4],[0],[4]]) # Actions
action_n = 3
assert(actions.shape[0] == action_n)
problem = make_hillcar_problem(step_len,
n_steps,
damp,
jitter,
discount,
bounds,
cost_radius,
actions) # Just needs the action boundaries
# Generate MDP
(mdp,disc) = make_uniform_mdp(problem,disc_n,action_n)
assert(np.all(mdp.actions == actions)) # Should be the same
#add_drain(mdp,disc,np.zeros(2),0)
return (mdp,disc,problem)
def solve_mdp_with_kojima(mdp):
# Solve
start = time.time()
(p,d) = solve_with_kojima(mdp,1e-8,1000,1e-8,1e-6)
print 'Kojima ran for:', time.time() - start, 's'
return (p,d)
if __name__ == '__main__':
####################################################
# Build the MDP and discretizer
(low_mdp,low_disc,problem) = build_problem(low_dim)
(ref_mdp,ref_disc,_) = build_problem(ref_dim)
####################################################
# Solve, initially, using Kojima
if BUILD_MODE == 'build':
# Build / load
(ref_p,_) = solve_mdp_with_kojima(ref_mdp)
ref_sol = block_solution(ref_mdp,ref_p)
(low_p,_) = solve_mdp_with_kojima(low_mdp)
low_sol = block_solution(low_mdp,low_p)
np.save('low_sol.npy', low_sol)
np.save('ref_sol.npy', ref_sol)
else:
assert(BUILD_MODE == 'load')
low_sol = np.load('low_sol.npy')
ref_sol = np.load('ref_sol.npy')
ref_v = ref_sol[:,0]
ref_v_fn = InterpolatedFunction(ref_disc,ref_v)
ref_p = ref_sol.reshape(-1,order='F')
low_v = low_sol[:,0]
"""
img = reshape_full(ref_v,ref_disc)
imshow(img)
plt.show()
"""
####################################################
# Form the Fourier projection (both value and flow)
B = get_basis_from_solution(ref_mdp,ref_disc,ref_sol,17*17)
(N,K) = B.shape
print 'Projected basis shape', B.shape
assert((N,) == ref_p.shape)
proj_p = B.dot(B.T.dot(ref_p))
proj_sol = block_solution(ref_mdp,proj_p)
proj_v = proj_sol[:,0]
#####################################################
# Build the start states
batched_start_states = create_start_states(STATES_PER_BATCH,
problem,
WORKERS * BATCHES_PER_WORKER)
start_states = np.vstack(batched_start_states)
#start_states = np.array([[4.0,3.0]])
#batched_start_states = [start_states]
#####################################################
# Rollout
rollout_policy = HillcarPolicy(ref_mdp.actions)
(rollout_ret,sim) = get_returns(problem,
rollout_policy,
ref_v_fn,
start_states,
SIM_HORIZON)
print 'Rollout policy:', np.percentile(rollout_ret,[25,50,75])
np.save(SAVE_FILE + 'rollout',rollout_ret)
"""
(N,d,T) = sim.states.shape
for i in xrange(N):
plt.plot(sim.states[i,0,:],
sim.states[i,1,:],
'x-k',alpha=0.5)
plt.show()
"""
#####################################################
# Run Q-policy with coarse grid values
(q_ret,sim) = get_q_returns(problem,
low_mdp,
low_disc,
low_v,
ref_v_fn,
start_states,
SIM_HORIZON)
print 'Coarse Q-policy:',np.percentile(q_ret,[25,50,75])
np.save(SAVE_FILE + 'q_low',q_ret)
"""
(N,d,T) = sim.states.shape
for i in xrange(N):
plt.plot(sim.states[i,0,:],
sim.states[i,1,:],
'x-k',alpha=0.5)
plt.show()
"""
#####################################################
# Run Q-policy with fine grid values
q_ret,_ = get_q_returns(problem,
ref_mdp,
ref_disc,
ref_v,
ref_v_fn,
start_states,
SIM_HORIZON)
print 'Fine Q-policy:',np.percentile(q_ret,[25,50,75])
np.save(SAVE_FILE + 'q_ref',q_ret)
"""
(N,d,T) = sim.states.shape
for i in xrange(N):
plt.plot(sim.states[i,0,:],
sim.states[i,1,:],
'x-k',alpha=0.5)
plt.show()
"""
#####################################################
# Run Q-policy with fourier values
q_ret,_ = get_q_returns(problem,
ref_mdp,
ref_disc,
proj_v,
ref_v_fn,
start_states,
SIM_HORIZON)
print 'Fourier Q-policy:',np.percentile(q_ret,[25,50,75])
np.save(SAVE_FILE + 'q_proj',q_ret)
####################################################
# MCTS with Coarse Q
# Implement the hillcar rollout policy in C++
for budget in BUDGETS:
mcts_params = MCTSParams(budget)
mcts_params.action_select_mode = ACTION_Q
fileroot = ROOT + '/hillcar.mcts'
mcts_ret = get_mcts_returns(DRIVER,
fileroot,
problem,
low_mdp,
low_disc,
low_sol,
ref_disc,
ref_v,
mcts_params,
batched_start_states,
SIM_HORIZON,
WORKERS)
print 'MCTS Coarse policy ({0}):'.format(budget),\
np.percentile(mcts_ret,[25,50,75])
np.save(SAVE_FILE + 'mcts_low_{0}'.format(budget), mcts_ret)
####################################################
# MCTS with Projected Q
for budget in BUDGETS:
mcts_params = MCTSParams(budget)
mcts_params.action_select_mode = ACTION_Q
fileroot = ROOT + '/hillcar.mcts'
mcts_ret = get_mcts_returns(DRIVER,
fileroot,
problem,
ref_mdp,
ref_disc,
proj_sol,
ref_disc,
ref_v,
mcts_params,
batched_start_states,
SIM_HORIZON,
WORKERS)
print 'MCTS projected policy ({0}):'.format(budget),\
np.percentile(mcts_ret,[25,50,75])
np.save(SAVE_FILE + 'mcts_proj_{0}'.format(budget), mcts_ret)
####################################################
# MCTS with No Q pessimistic
noq_sol = np.array(low_sol)
noq_sol[:,0] = 1.0/ (1.0 - problem.discount) # HOPELESS.
for budget in BUDGETS:
mcts_params = MCTSParams(budget)
mcts_params.action_select_mode = ACTION_Q
fileroot = ROOT + '/hillcar.mcts'
mcts_ret = get_mcts_returns(DRIVER,
fileroot,
problem,
low_mdp,
low_disc,
noq_sol,
ref_disc,
ref_v,
mcts_params,
batched_start_states,
SIM_HORIZON,
WORKERS)
print 'MCTS Coarse no Q policy ({0}):'.format(budget),\
np.percentile(mcts_ret,[25,50,75])
np.save(SAVE_FILE + 'mcts_noq_pes_{0}'.format(budget), mcts_ret)
####################################################
# MCTS with No Q Optimistic
noq_sol = np.array(low_sol)
noq_sol[:,0] = 0 # Hey! We're done!
for budget in BUDGETS:
mcts_params = MCTSParams(budget)
mcts_params.action_select_mode = ACTION_Q
fileroot = ROOT + '/hillcar.mcts'
mcts_ret = get_mcts_returns(DRIVER,
fileroot,
problem,
low_mdp,
low_disc,
noq_sol,
ref_disc,
ref_v,
mcts_params,
batched_start_states,
SIM_HORIZON,
WORKERS)
print 'MCTS Coarse no Q policy ({0}):'.format(budget),\
np.percentile(mcts_ret,[25,50,75])
np.save(SAVE_FILE + 'mcts_noq_opt_{0}'.format(budget), mcts_ret)
####################################################
# MCTS with No Flow
noflow_sol = np.array(low_sol)
noflow_sol[:,1:] = 1.0
for budget in BUDGETS:
mcts_params = MCTSParams(budget)
mcts_params.action_select_mode = ACTION_Q
fileroot = ROOT + '/hillcar.mcts'
mcts_ret = get_mcts_returns(DRIVER,
fileroot,
problem,
low_mdp,
low_disc,
noflow_sol,
ref_disc,
ref_v,
mcts_params,
batched_start_states,
SIM_HORIZON,
WORKERS)
print 'MCTS Coarse no flow policy ({0}):'.format(budget),\
np.percentile(mcts_ret,[25,50,75])
np.save(SAVE_FILE + 'mcts_noflow_{0}'.format(budget), mcts_ret)
####################################################
# MCTS with no MDP information
neither_sol = np.empty(low_sol.shape)
neither_sol[:,0] = 1.0/ (1.0 - problem.discount)
neither_sol[:,1:] = 1.0
for budget in BUDGETS:
mcts_params = MCTSParams(budget)
mcts_params.action_select_mode = ACTION_Q
fileroot = ROOT + '/hillcar.mcts'
mcts_ret = get_mcts_returns(DRIVER,
fileroot,
problem,
low_mdp,
low_disc,
neither_sol,
ref_disc,
ref_v,
mcts_params,
batched_start_states,
SIM_HORIZON,
WORKERS)
print 'MCTS Coarse no mpd information policy ({0}):'.format(budget),\
np.percentile(mcts_ret,[25,50,75])
np.save(SAVE_FILE + 'mcts_noflow_{0}'.format(budget), mcts_ret)