def __init__(self, env, actor, critic, DISCOUNT_FACTOR, MINIBATCH_SIZE,
REPLAY_MEMORY_SIZE, TARGET_DISCOUNT, continue_execution,
MEMORIES):
# Environment details
self.env = env
self.actor = actor
self.critic = critic
self.MINIBATCH_SIZE = MINIBATCH_SIZE
self.DISCOUNT = DISCOUNT_FACTOR
self.TARGET_DISCOUNT = TARGET_DISCOUNT
self.bg_noise = None
self.action_dim = self.env.action_space.shape[0]
# Replay memory to store experiences of the model with the environment
# self.replay_memory = deque(maxlen=REPLAY_MEMORY_SIZE)
if continue_execution:
self.replay_memory = memory.Memory(REPLAY_MEMORY_SIZE,
load=continue_execution,
memories=MEMORIES)
else:
self.replay_memory = memory.Memory(REPLAY_MEMORY_SIZE)
_state += list(states[i]) + list(actions[i])
_state += list(states[len(states) - 1])
return states, actions, np.asarray(tuple(_state))
#start iterating from 'current epoch'
for episode in xrange(CURRENT_EPISODE + 1, EPISODES + 1, 1):
done = False
first_state = env.reset()
first_action = np.array([0, 0, 0])
states = [first_state, first_state, first_state]
actions = [first_action, first_action]
states, actions, cur_state = make_state(states, actions, first_state,
first_action)
action_memory = memory.Memory(STEPS)
episode_reward = 0
episode_step = 0
new_episode = True
while not done:
action, action_step = actor_critic.act(cur_state, new_episode,
GREEDY_RATE)
_next_state, reward, done, _ = env.step(action_step)
states, actions, next_state = make_state(states, actions,
_next_state, action)
episode_reward += reward
action_memory.addMemory(cur_state, action, reward, next_state,
done)
