def init_transitions(self, horizon):
neutral_tmat = generate_transitions(self.states, self.actions,
self.true_tran_func)
realvals, _ = generate_policy(horizon, self.max_time,
neutral_tmat, self.rmat)
for i, (aname, gamma) in enumerate(self.attitudes):
transition_func = self.make_transition_func(aname, gamma, realvals)
self.transitions[aname] = generate_transitions(self.states, self.actions,
f, self.max_time)
fname = self.dirname + "transitions_{0}_{1}_h{2}.p".format(i, aname, horizon)
self.transitions[aname].dump(fname)
def run_sims(dirname, horizons, attitudes, model_type, model, num_episodes):
max_time = max(horizons)
time_based = False
if model_type == "valopt":
time_based = True
# output setup
csvout = open(dirname + "results_{0}_data.csv".format(num_episodes), 'w+')
writer = csv.writer(csvout)
writer.writerow(["Horizon", "AvgEarnings", "StdErr", "AttitudeIndex", "AttitudeName"])
readableout = open(dirname + "results_{0}_readable.txt".format(num_episodes), 'w+')
verboseout = open(dirname + "results_{0}_verbose.txt".format(num_episodes), 'w+')
print_readverb = make_multiprint([readableout, verboseout])
# variables only used for constvar
# TODO: allow for >2 actions?
num_actions = 2
new_attitudes = []
policies = []
if model_type == "constvar":
for action_idx in range(num_actions):
policies.append(const_policy(action_idx, max_time, len(model.states)))
results = np.zeros((len(attitudes) * num_actions, len(horizons), 3))
else:
results = np.zeros((len(attitudes), len(horizons), 3))
for horizon_idx, horizon in enumerate(horizons):
print_readverb("Horizon: {}".format(horizon))
if model_type == "constvar":
"""In the `constvar` model, we vary the true transition matrix,
rather than the policy, based on beliefs. We take a constant policy
(one for each action)."""
for attitude_idx, attitude in enumerate(attitudes):
# compute results
tmat = model.transitions[attitude]
rmat = model.rewards[attitude]
for action_idx in range(num_actions):
new_att_idx = (attitude_idx * num_actions) + action_idx
new_attitude = attitude + str(action_idx)
if len(new_attitudes) < len(attitudes) * num_actions:
new_attitudes.append(new_attitude)
totals, mean, stderr = episode_stats(policies[action_idx], num_episodes,
max_time, tmat, rmat, time_based)
results[new_att_idx][horizon_idx] = [horizon, mean, stderr]
# output to files
writer.writerow([horizon, mean, stderr, new_att_idx, new_attitude])
print_readverb("{0}: mean = {1}, stderr = {2}".format(new_attitude, mean, stderr))
print("Totals: {0}".format(totals), file=verboseout)
print("Policy for {0}: {1}".format(new_attitude, policies[action_idx]), file=verboseout)
else:
if model_type == "valopt":
model.init_transitions(horizon)
# note: "neutral" must be an attitude
neutral_tmat = model.transitions['neutral']
for attitude_idx, attitude in enumerate(attitudes):
rmat = model.rewards[attitude]
if model_type == "const":
# TODO: allow const to have something other than 0
policy = const_policy(0, max_time, rmat.shape[1])
else:
tmat = model.transitions[attitude]
_, policy = generate_policy(horizon, max_time, tmat, rmat, time_based)
totals, mean, stderr = episode_stats(policy, num_episodes, max_time,
neutral_tmat, rmat, time_based)
results[attitude_idx][horizon_idx] = [horizon, mean, stderr]
# output to files
writer.writerow([horizon, mean, stderr, attitude_idx, attitude])
print_readverb("{0}: mean = {1}, stderr = {2}".format(attitude, mean, stderr))
print("Totals: {0}".format(totals), file=verboseout)
print("Policy for {0}: {1}".format(attitude, policy), file=verboseout)
polfname = "policy_{0}_{1}_{2}_{3}.p".format(num_episodes, horizon,
attitude_idx, attitude)
cPickle.dump((attitude, policy), open(dirname + polfname, 'w+'))
if len(new_attitudes) > 0:
attitudes = new_attitudes
return results, attitudes
