def run_episode(environment: gym.Env, agent: DQNAgent, render: bool,
max_length: int):
"""
Run one episode in the given environment with the agent.
Arguments:
environment {`gym.Env`} -- Environment representing the Markov Decision Process
agent {`DQNAgent`} -- Reinforcment Learning agent that acts in the envíronment
render {`bool`} -- Whether the frames of the episode should be rendered on the screen
max_length {`int`} -- Maximum number of steps before the episode is terminated
Returns:
`float` -- Cumulated reward that the agent received during the episode
"""
episode_reward = 0
state = environment.reset()
for _ in range(max_length):
if render:
environment.render()
action = agent.act(state)
next_state, reward, terminal, _ = environment.step(action)
agent.observe(
Transition(state, action, reward,
None if terminal else next_state))
episode_reward += reward
if terminal:
break
else:
state = next_state
return episode_reward
train_interval=1,
batch_size=32,
memory_interval=1,
observation=obs,
input_shape=[len(obs)],
training=True,
policy=policy)
agent.compile()
result = []
for episode in range(500): # 1000エピソード回す
agent.reset()
observation = env.reset() # 環境の初期化
# observation, _, _, _ = env.step(env.action_space.sample())
observation = deepcopy(observation)
agent.observe(observation)
for t in range(250): # n回試行する
# env.render() # 表示
action = agent.act()
observation, reward, done, info = env.step(
action) # アクションを実行した結果の状態、報酬、ゲームをクリアしたかどうか、その他の情報を返す
observation = deepcopy(observation)
agent.observe(observation, reward, done)
if done:
break
# test
agent.training = False
observation = env.reset() # 環境の初期化
agent.observe(observation)
for t in range(250):