Function main

reinforcement_learning/actor_critic.py:138–177

()

def main():
    running_reward = 10

    # run infinitely many episodes
    for i_episode in count(1):

        # reset environment and episode reward
        state, _ = env.reset()
        ep_reward = 0

        # for each episode, only run 9999 steps so that we don't
        # infinite loop while learning
        for t in range(1, 10000):

            # select action from policy
            action = select_action(state)

            # take the action
            state, reward, terminated, truncated, _ = env.step(action)

            model.rewards.append(reward)
            ep_reward += reward
            if terminated or truncated:
                break

        # update cumulative reward
        running_reward = 0.05 * ep_reward + (1 - 0.05) * running_reward

        # perform backprop
        finish_episode()

        # log results
        if i_episode % args.log_interval == 0:
            print(f'Episode {i_episode}\tLast reward: {ep_reward:.2f}\tAverage reward: {running_reward:.2f}')

        # check if we have "solved" the cart pole problem
        if running_reward > env.spec.reward_threshold:
            print(f"Solved! Running reward is now {running_reward} and "
                  f"the last episode runs to {t} time steps!")
            break


if __name__ == '__main__':