DQNAgent

agent.py

class DQNAgent:
    
    """ ==== 中略 ==== """
    
    def play(self, episodes):

        total_rewards = []

        for n in range(episodes):

            self.epsilon = 1.0 - min(0.95, self.global_steps * 0.95 / 500)

            total_reward = self.play_episode()

            total_rewards.append(total_reward)

            print(f"Episode {n}: {total_reward}")
            print(f"Current experiences {len(self.experiences)}")
            print(f"Current epsilon {self.epsilon}")
            print()

        return total_rewards

    def play_episode(self):

        total_reward = 0

        episode_steps = 0

        done = False

        state = self.env.reset()

        while not done:

            action = self.sample_action(state)

            next_state, reward, done, info = self.env.step(action)

            total_reward += reward

            exp = Experience(state, action, reward, next_state, done)

            self.experiences.append(exp)

            state = next_state

            episode_steps += 1

            self.global_steps += 1
            
            self.update_qnetwork()
            
            #: target_networkの更新
            if self.global_steps % 250 == 0:
                self.target_network.set_weights(self.q_network.get_weights())

        return total_reward