def train(model_path='models/model.h5',
opponent_policy=random_choice,
num_episodes=1000,
agent_params={},
**kwargs):
stats = statistics.default_stats()
plt_data = statistics.plot_stats(stats, data=None)
agent = DQNAgent(**agent_params)
for episode in range(num_episodes):
print('Episode {}/{}'.format(episode, num_episodes))
env = Environment(opponent_policy=opponent_policy,
agent_color=board.RED,
agent_first_turn=True)
done = False
episode_length = 0
while not done:
state = env.get_state()
action = agent.act_epsilon_greedy(state)
next_state, reward, event = env.step(action)
done = event != board.EVENT_IN_GAME
agent.remember(state, action, reward, next_state, done)
agent.replay(stats=stats)
if event == board.EVENT_WIN:
print('Won Game!')
episode_length += 1
stats['episode_results'].append(event)
stats['episode_lengths'].append(episode_length)
plt_data = statistics.plot_stats(stats, data=plt_data)
plt.pause(0.0001)
if episode % 100 == 0:
agent.save(model_path)
agent.save(model_path)
saved_params = {'agent_params': agent_params, 'num_episodes': num_episodes}
statistics.save_stats(
stats, saved_params,
"stats/stats-{}.json".format(time.strftime("%Y%m%d-%H%M%S")))
statistics.plot_stats(stats, data=plt_data)
plt.show()
episode_reward = 0.0
episode_count = 0
obs = env.reset()
pb = tqdm(range(-params.memory_initial, params.max_steps))
plotter = VisdomLinePlotter()
reward_history = deque(maxlen=100)
for i in pb:
if args.render:
env.render()
idx = memory.store_obs(obs)
state = memory.get_stacked_obs(idx)
action = agent.act_epsilon_greedy(state, eps_schedule.get(i))
next_obs, reward, done, _ = env.step(action)
episode_reward += reward
memory.store_effect(idx, action, np.sign(reward), done)
if done:
next_obs = env.reset()
episode_count += 1
reward_history.append(episode_reward)
pb.set_description(
f"episode: {episode_count}, reward: {episode_reward}, eps: {eps_schedule.get(i)*100:.2f}%"
)
plotter.plot('episode reward', 'episode return', "Episode Return",
episode_count, episode_reward)
plotter.plot('episode reward', 'average return', "Episode Return",