model.fit(states, actions)
print ("Model trained in %.03f sec." % tm.time)
failed = 0
sampled = np.zeros((ntrials, d, n))
ncases = model._cb_t._counter - 1
for j in xrange(ntrials):
with Timer() as tm:
# Test model
iter_ = 0
nfails = 0
while True:
try:
sampled[j, :, 0] = model.sample()
for iter_, a in enumerate(actions.T):
# sample next state resulting from executing action `a` in state `state`
next_state = model.sample(MDPState(sampled[j, :, iter_]), a)
if next_state is None:
raise TypeError
sampled[j, :, iter_ + 1] = next_state
except:
# plot_sampled(obs[:j], sampled[i, j, :, :iter_])
sampled[j, :].fill(0)
nfails += 1
print "{0}:{1} Failed to infer next state distribution at step {2}.".format(
j, nfails, iter_
n = obs.shape[0]
action_error = -np.inf * np.ones(n)
delta_error = -np.inf * np.ones(n)
for i, states in enumerate(obs):
# Train CASML's case base and hmm with states and actions
model.fit(states, actions)
# Test model
cntr = 0
iter_ = 0
while cntr < 10:
sampled = None
try:
sampled = np.array([model.sample()]).T
for iter_, a in enumerate(actions.T):
# sample next state resulting from executing action `a` in state `state`
next_state = model.sample(MDPState(sampled[:, -1]), a)[:, np.newaxis]
sampled = np.hstack([sampled, next_state])
except:
print "{0}:{1} Failed to infer next state distribution at step {2}.".format(i + 1, cntr + 1, iter_ + 1)
# plot_sampled(obs[0:i+1], sampled)
cntr += 1
continue
break
if cntr < 10:
action_error[i] = evaluate_action(actions, obs[0:i + 1], sampled, plot=True)[0]
delta_error[i] = evaluate_delta(obs[0:i + 1], sampled, plot=True)
