epsilon = 10**(-12)
samples = collect_samples(sim)
if len(samples) < 5000:
samples += collect_samples(sim)
discount = .8
# construct a graph from the samples
graph = pvf.construct_graph(samples, Simulator.states)
basis = pvf.create_basis_function(graph, Simulator.states,
Simulator.actions, k)
policy = initialize_policy(0.0, discount, basis)
final_policy, all_policies = lspi.lspi(maxiter, epsilon,
samples, policy)
value_policy = initialize_value_function_policy(sim)
plt.figure()
plt.subplot(2,2,1)
approxV = display_qvalues(sim, final_policy)
plt.title('Estimated Value Function')
plt.subplot(2,2,2)
display_qvalues(sim, final_policy, dim=1)
plt.title('Estimated Value Function')
#lt.subplot(1,2,2)
#isplay_policy(sim, final_policy)
#lt.show()
if (k, episode) not in data:
data[(k, episode)] = []
samples = rooms.collect_samples(sim, maxepisodes=episode, maxsteps=max_steps)
graph = pvf.construct_graph(samples, sim.states)
try:
basis = pvf.create_basis_function(graph, sim.states,
sim.actions, k)
except:
print "Couldn't compute basis function for this data"
continue
policy = rooms.initialize_policy(0.0, discount, basis)
final_policy = lspi.lspi(maxiter, epsilon,
samples, policy)[0]
for n in range(num_tries):
execution_data = rooms.test_execution(sim, final_policy, maxsteps=max_steps)
data[(k, episode)].append(execution_data)
for episode in range(start_episode, end_episode+1, step_episode):
for k in range(start_k, end_k+1, step_k):
total_steps = 0
data_list = data.get((k, episode), [])
for data_point in data_list:
total_steps += data_point[2]
if k not in final_data: