def agent_loop(dictionary, lock1, lock2):
random.seed()
environment = GridWorldModel()
agent = Agent(environment)
agent.Q = dictionary[Q_SHARED_KEY] # initialize with shared Q
while environment.step_count < MAX_STEPS_PER_AGENT:
environment.reset()
agent.state = environment.get_start_state()
while True:
agent.act()
if environment.step_count % ASYNC_UPDATE_INTERVAL == 0 or environment.is_terminal_state(
):
lock1.acquire()
q = dictionary[Q_SHARED_KEY]
# Need to write it back, otherwise the proxy won't pick up the changes.
dictionary[Q_SHARED_KEY] = np.add(q, agent.dQ)
lock1.release()
agent.dQ = np.zeros((GridWorldModel.get_number_of_states(),
GridWorldModel.get_number_of_actions()),
dtype=float)
if environment.is_terminal_state():
break
lock2.acquire()
combined_rewards = dictionary[REWARDS_KEY]
agents_rewards = np.array(agent.rewards)
# ...same here
dictionary[REWARDS_KEY] = np.add(combined_rewards,
agents_rewards[:MAX_STEPS_PER_AGENT])
lock2.release()
