def arrange_match(self):
self.episode_count = self.episode_count + 1
brain_name = self.env.brain_names[0]
env_info = self.env.reset(train_mode=True)[brain_name]
state = env_info.vector_observations
episode_reward1 = 0
episode_reward2 = 0
for step in range(self.max_steps):
actions = [self.agent.get_action(state[0]), self.agent.get_action(state[1])]
env_info = self.env.step(actions)[brain_name]
reward = env_info.rewards
next_state = env_info.vector_observations
done = env_info.local_done
self.agent.learn_experience(buffer.Experience(state[0], actions[0], reward[0], next_state[0], done[0]))
self.agent.learn_experience(buffer.Experience(state[1], actions[1], reward[1], next_state[1], done[1]))
episode_reward1 += reward[0]
episode_reward2 += reward[1]
if done[0] or done[1] or step == self.max_steps-1:
break
state = next_state
return max(episode_reward1, episode_reward2), step
def learn_episode(self, max_steps):
brain_name = self.env.brain_names[0]
env_info = self.env.reset(train_mode=True)[brain_name]
state = env_info.vector_observations[0]
episode_reward = 0
for step in range(max_steps):
action = self.get_action(state)
env_info = self.env.step(action)[brain_name]
reward = env_info.rewards[0]
next_state = env_info.vector_observations[0]
done = env_info.local_done[0]
self.replay_buffer.add(buffer.Experience(state, action, reward, next_state, done))
episode_reward += reward
if self.replay_buffer.ready_to_sample():
self.update()
if done or step == max_steps-1:
break
state = next_state
return episode_reward
def play_episode(self):
brain_name = self.env.brain_names[0]
env_info = self.env.reset(train_mode=True)[brain_name]
state = env_info.vector_observations
episode_reward1 = 0
episode_reward2 = 0
for step in range(self.max_steps):
actions = self.agent.get_actions(state)
env_info = self.env.step(actions)[brain_name]
reward = env_info.rewards
next_state = env_info.vector_observations
done = env_info.local_done
self.agent.learn_experience(buffer.Experience(state, actions, reward, next_state, done))
episode_reward1 += reward[0]
episode_reward2 += reward[1]
if done[0] or done[1] or step == self.max_steps-1:
break
state = next_state
return max(episode_reward1, episode_reward2), step
def arrange_match(self):
self.episode_count = self.episode_count + 1
if self.steps_till_freeze < 0:
self.steps_till_freeze = self.freeze_steps
self.frozen_agents[self.next_freeze_agent].copy_and_freeze(self.agent)
self.next_freeze_agent = (self.next_freeze_agent + 1) % len(self.frozen_agents)
brain_name = self.env.brain_names[0]
env_info = self.env.reset(train_mode=True)[brain_name]
state = env_info.vector_observations
episode_reward1 = 0
episode_reward2 = 0
if random.random() < self.self_play_probability:
opponent_agent = self.agent
else:
opponent_agent = random.choice(self.frozen_agents)
for step in range(self.max_steps):
actions = [self.agent.get_action(state[0]), opponent_agent.get_action(state[1])]
env_infos = self.env.step(actions)
reward = env_infos[brain_name].rewards
next_state = env_infos[brain_name].vector_observations
done = env_infos[brain_name].local_done
self.agent.learn_experience(buffer.Experience(state[0], actions[0], reward[0], next_state[0], done[0]))
self.agent.learn_experience(buffer.Experience(state[1], actions[1], reward[1], next_state[1], done[1]))
episode_reward1 += reward[0]
episode_reward2 += reward[1]
if done[0] or done[1] or step == self.max_steps-1:
break
self.steps_till_freeze -= 1
state = next_state
return max(episode_reward1, episode_reward2), step