как реализовать это без использования библиотеки ptan? - PullRequest
Новая
       14

как реализовать это без использования библиотеки ptan?

0 голосов
/ 08 июля 2019

Я изучаю RL из книги Максима Лапана Deep_Reinforcement_Learning_Hands-On. Вот их код для алгоритма REINFORCE. https://github.com/PacktPublishing/Deep-Reinforcement-Learning-Hands-On/blob/master/Chapter09/02_cartpole_reinforce.py Но хочу реализовать этот код без использования библиотеки ptan, так как я не понимаю, что она делает.

Я попытался сделать код ниже 

    env = gym.make("CartPole-v0")
    writer = SummaryWriter(comment="-cartpole-reinforce")

    net = PGN(env.observation_space.shape[0], env.action_space.n)
    print(net)


    optimizer = optim.Adam(net.parameters(), lr=LEARNING_RATE)

    total_rewards = []
    step_idx = 0
    done_episodes = 0

    batch_episodes = 0
    batch_states, batch_actions, batch_qvals = [], [], []
    cur_rewards = []
    state=env.reset()
    for step_idx in range(100):

        batch_states.append(state)
        state=torch.Tensor(state)
        action=int(net(state).detach().max(0)[1])
        batch_actions.append(action)
        state,r,done,_=env.step(action)
        cur_rewards.append(r)

        if done:
            batch_qvals.extend(calc_qvals(cur_rewards))
            reward=sum(cur_rewards)
            cur_rewards.clear()
            batch_episodes += 1
            state=env.reset()

            # handle new rewards

        if done:
            done_episodes += 1
            total_rewards.append(reward)
            mean_rewards = float(np.mean(total_rewards[-100:]))
            print("%d: reward: %6.2f, mean_100: %6.2f, episodes: %d" % (
                step_idx, reward, mean_rewards, done_episodes))
            writer.add_scalar("reward", reward, step_idx)
            writer.add_scalar("reward_100", mean_rewards, step_idx)
            writer.add_scalar("episodes", done_episodes, step_idx)
            if mean_rewards > 195:
                print("Solved in %d steps and %d episodes!" % (step_idx, done_episodes))
                break

        if batch_episodes < EPISODES_TO_TRAIN:
            continue

        optimizer.zero_grad()
        states_v = torch.FloatTensor(batch_states)
        batch_actions_t = torch.LongTensor(batch_actions)
        batch_qvals_v = torch.FloatTensor(batch_qvals)

        logits_v = net(states_v)
        log_prob_v = F.log_softmax(logits_v, dim=1)
        log_prob_actions_v = batch_qvals_v * log_prob_v[range(len(batch_states)), batch_actions_t]
        loss_v = -log_prob_actions_v.mean()

        loss_v.backward()
        optimizer.step()

        batch_episodes = 0
        batch_states.clear()
        batch_actions.clear()
        batch_qvals.clear()

    writer.close()
``
But this doesn't converge at all, the reward is never more than 10
...