def testPolicy(): num_actions = 6 start_value = 1 end_value = 0.1 num_steps = 1000 policy1 = GreedyPolicy() policy2 = LinearDecayGreedyEpsilonPolicy(num_actions, start_value, end_value, num_steps) q_values = np.array([1.0, 1.3, 1.2, 1.5, 1.1, 1.4]) assert(policy1.select_action(q_values) == 3) assert(policy2.epsilon == 1) policy2.select_action(q_values) assert(np.isclose(policy2.epsilon, 0.9991)) policy2.select_action(q_values) policy2.select_action(q_values) assert(np.isclose(policy2.epsilon, 0.9973)) for i in range(num_steps): policy2.select_action(q_values) assert(np.isclose(policy2.epsilon, end_value)) policy2.select_action(q_values) assert(np.isclose(policy2.epsilon, end_value))
history_store.reset()
state = atari_processor.state_for_nn(observation)
history_store.add_history(state)
nn_tmp = history_store.get_history()
nn_input = np.zeros((1, IMAGE_SIZE[0], IMAGE_SIZE[1], HISTORY_LENGTH),
dtype=float)
nn_input[0, :] = nn_tmp
episode_interaction_cnt = 0
flag_first = 0
done = False
while done == False:
# Interact with environment and store into memory.
q_values = model_online.predict(nn_input)
action = greedy_epsilon_linear_decay_selector.select_action(q_values)
observation, reward, done, info = env.step(action)
if flag_first == 0:
info_prev = info
episode_end == False
flag_first = 1
else:
if info != info_prev:
episode_end = True
else:
episode_end = False
info_prev = info
reward = atari_processor.process_reward(reward)
state_next = atari_processor.state_for_nn(observation)
history_store.add_history(state_next)