def _update_table_sample():
def _update_table_model():
def _lookup_table():
if __name__ == "__main__":
maze = SimpleMaze()
agent = QAgent(actions=maze.ACTIONS, alpha=0.5, gamma=0.5, explore_strategy='epsilon', epsilon=0.1)
# logging
path = deque() # path in this episode
episode_reward_rates = []
num_episodes = 0
cum_reward = 0
cum_steps = 0
# repeatedly run episodes
while True:
# initialization
maze.reset()
agent.reset(foget_table=False)
action, _ = agent.observe_and_act(observation=None, last_reward=None) # get and random action
path.clear()
episode_reward = 0
episode_steps = 0
# interact and reinforce repeatedly
while not maze.isfinished():
new_observation, reward = maze.interact(action)
action, _ = agent.observe_and_act(observation=new_observation, last_reward=reward)
path.append(new_observation)
episode_reward += reward
episode_steps += 1
print len(path),
cum_steps += episode_steps
cum_reward += episode_reward
num_episodes += 1
episode_reward_rates.append(episode_reward / episode_steps)
if num_episodes % 100 == 0:
print num_episodes, len(agent.q_table), cum_reward, cum_steps, 1.0 * cum_reward / cum_steps#, path
cum_reward = 0
cum_steps = 0
win = 50
buffer_idx = np.random.randint(0, self.NUM_BUFFERS, (self.BATCH_SIZE,))
return (self.buffer_old_state[buffer_idx, sample_idx, :],
self.buffer_action[buffer_idx, sample_idx],
self.buffer_reward[buffer_idx, sample_idx],
self.buffer_new_state[buffer_idx, sample_idx, :])
def isfilled(self):
return all(self.filled)
def reset(self):
self.top = [-1]*self.NUM_BUFFERS
self.filled = [False]*self.NUM_BUFFERS
if __name__ == '__main__':
maze = SimpleMaze()
agent = QAgentNN(dim_state=(1, 1, 2), range_state=((((0, 3),(0, 4)),),), actions=maze.ACTIONS,
learning_rate=0.01,
reward_scaling=100.0, reward_scaling_update='adaptive', rs_period=2,
batch_size=100, update_period=10,
freeze_period=2, memory_size=1000,
alpha=0.5, gamma=0.5, explore_strategy='epsilon', epsilon=0.02, verbose=2)
print "Maze and agent initialized!"
# logging
path = deque() # path in this episode
episode_reward_rates = []
num_episodes = 0
cum_reward = 0
cum_steps = 0
