Esempio n. 1
0
class COMA:
    def __init__(self, args):
        self.n_actions = args.n_actions
        self.n_agents = args.n_agents
        self.state_shape = args.state_shape
        self.obs_shape = args.obs_shape
        actor_input_shape = self.obs_shape  # actor网络输入的维度,和vdn、qmix的rnn输入维度一样,使用同一个网络结构
        critic_input_shape = self._get_critic_input_shape()  # critic网络输入的维度
        # 根据参数决定RNN的输入维度
        if args.last_action:
            actor_input_shape += self.n_actions
        if args.reuse_network:
            actor_input_shape += self.n_agents
        self.args = args

        # 神经网络
        # 每个agent选动作的网络,输出当前agent所有动作对应的概率,用该概率选动作的时候还需要用softmax再运算一次。
        if self.args.alg == 'coma':
            print('Init alg coma')
            self.eval_rnn = RNN(actor_input_shape, args)
        elif self.args.alg == 'coma+commnet':
            print('Init alg coma+commnet')
            self.eval_rnn = CommNet(actor_input_shape, args)
        elif self.args.alg == 'coma+g2anet':
            print('Init alg coma+g2anet')
            self.eval_rnn = G2ANet(actor_input_shape, args)
        else:
            raise Exception("No such algorithm")

        # 得到当前agent的所有可执行动作对应的联合Q值,得到之后需要用该Q值和actor网络输出的概率计算advantage
        self.eval_critic = ComaCritic(critic_input_shape, self.args)
        self.target_critic = ComaCritic(critic_input_shape, self.args)

        if self.args.cuda:
            self.eval_rnn.cuda()
            self.eval_critic.cuda()
            self.target_critic.cuda()

        self.model_dir = args.model_dir + '/' + args.alg + '/' + args.map
        # 如果存在模型则加载模型
        # if os.path.exists(self.model_dir + '/rnn_params.pkl'):
        #     path_rnn = self.model_dir + '/rnn_params.pkl'
        #     path_coma = self.model_dir + '/critic_params.pkl'
        #     self.eval_rnn.load_state_dict(torch.load(path_rnn))
        #     self.eval_critic.load_state_dict(torch.load(path_coma))
        #     print('Successfully load the model: {} and {}'.format(path_rnn, path_coma))

        # 让target_net和eval_net的网络参数相同
        self.target_critic.load_state_dict(self.eval_critic.state_dict())

        self.rnn_parameters = list(self.eval_rnn.parameters())
        self.critic_parameters = list(self.eval_critic.parameters())

        if args.optimizer == "RMS":
            self.critic_optimizer = torch.optim.RMSprop(self.critic_parameters,
                                                        lr=args.lr_critic)
            self.rnn_optimizer = torch.optim.RMSprop(self.rnn_parameters,
                                                     lr=args.lr_actor)
        self.args = args

        # 执行过程中,要为每个agent都维护一个eval_hidden
        # 学习过程中,要为每个episode的每个agent都维护一个eval_hidden
        self.eval_hidden = None

    def _get_critic_input_shape(self):
        # state
        input_shape = self.state_shape  # 48
        # obs
        input_shape += self.obs_shape  # 30
        # agent_id
        input_shape += self.n_agents  # 3
        # 所有agent的当前动作和上一个动作
        input_shape += self.n_actions * self.n_agents * 2  # 54

        return input_shape

    def learn(self, batch, max_episode_len, train_step,
              epsilon):  # train_step表示是第几次学习,用来控制更新target_net网络的参数
        episode_num = batch['o'].shape[0]
        self.init_hidden(episode_num)
        for key in batch.keys():  # 把batch里的数据转化成tensor
            if key == 'u':
                batch[key] = torch.tensor(batch[key], dtype=torch.long)
            else:
                batch[key] = torch.tensor(batch[key], dtype=torch.float32)
        u, r, avail_u, terminated = batch['u'], batch['r'], batch[
            'avail_u'], batch['terminated']
        mask = (1 - batch["padded"].float()).repeat(
            1, 1, self.n_agents)  # 用来把那些填充的经验的TD-error置0,从而不让它们影响到学习
        if self.args.cuda:
            u = u.cuda()
            mask = mask.cuda()
        # 根据经验计算每个agent的Q值,从而跟新Critic网络。然后计算各个动作执行的概率,从而计算advantage去更新Actor。
        q_values = self._train_critic(
            batch, max_episode_len,
            train_step)  # 训练critic网络,并且得到每个agent的所有动作的Q值
        action_prob = self._get_action_prob(batch, max_episode_len,
                                            epsilon)  # 每个agent的所有动作的概率

        q_taken = torch.gather(q_values, dim=3,
                               index=u).squeeze(3)  # 每个agent的选择的动作对应的Q值
        pi_taken = torch.gather(action_prob, dim=3,
                                index=u).squeeze(3)  # 每个agent的选择的动作对应的概率
        pi_taken[mask ==
                 0] = 1.0  # 因为要取对数,对于那些填充的经验,所有概率都为0,取了log就是负无穷了,所以让它们变成1
        log_pi_taken = torch.log(pi_taken)

        # 计算advantage
        baseline = (q_values * action_prob).sum(
            dim=3, keepdim=True).squeeze(3).detach()
        advantage = (q_taken - baseline).detach()
        loss = -((advantage * log_pi_taken) * mask).sum() / mask.sum()
        self.rnn_optimizer.zero_grad()
        loss.backward()
        torch.nn.utils.clip_grad_norm_(self.rnn_parameters,
                                       self.args.grad_norm_clip)
        self.rnn_optimizer.step()
        # print('Training: loss is', loss.item())
        # print('Training: critic params')
        # for params in self.eval_critic.named_parameters():
        #     print(params)
        # print('Training: actor params')
        # for params in self.eval_rnn.named_parameters():
        #     print(params)

    def _get_critic_inputs(self, batch, transition_idx, max_episode_len):
        # 取出所有episode上该transition_idx的经验
        obs, obs_next, s, s_next = batch['o'][:, transition_idx], batch['o_next'][:, transition_idx],\
                                   batch['s'][:, transition_idx], batch['s_next'][:, transition_idx]
        u_onehot = batch['u_onehot'][:, transition_idx]
        if transition_idx != max_episode_len - 1:
            u_onehot_next = batch['u_onehot'][:, transition_idx + 1]
        else:
            u_onehot_next = torch.zeros(*u_onehot.shape)
        # s和s_next是二维的,没有n_agents维度,因为所有agent的s一样。其他都是三维的,到时候不能拼接,所以要把s转化成三维的
        s = s.unsqueeze(1).expand(-1, self.n_agents, -1)
        s_next = s_next.unsqueeze(1).expand(-1, self.n_agents, -1)
        episode_num = obs.shape[0]
        # 因为coma的critic用到的是所有agent的动作,所以要把u_onehot最后一个维度上当前agent的动作变成所有agent的动作
        u_onehot = u_onehot.view(
            (episode_num, 1, -1)).repeat(1, self.n_agents, 1)
        u_onehot_next = u_onehot_next.view(
            (episode_num, 1, -1)).repeat(1, self.n_agents, 1)

        if transition_idx == 0:  # 如果是第一条经验,就让前一个动作为0向量
            u_onehot_last = torch.zeros_like(u_onehot)
        else:
            u_onehot_last = batch['u_onehot'][:, transition_idx - 1]
            u_onehot_last = u_onehot_last.view(
                (episode_num, 1, -1)).repeat(1, self.n_agents, 1)

        inputs, inputs_next = [], []
        # 添加状态
        inputs.append(s)
        inputs_next.append(s_next)
        # 添加obs
        inputs.append(obs)
        inputs_next.append(obs_next)
        # 添加所有agent的上一个动作
        inputs.append(u_onehot_last)
        inputs_next.append(u_onehot)

        # 添加当前动作
        '''
        因为coma对于当前动作,输入的是其他agent的当前动作,不输入当前agent的动作,为了方便起见,每次虽然输入当前agent的
        当前动作,但是将其置为0相量,也就相当于没有输入。
        '''
        action_mask = (1 - torch.eye(self.n_agents))  # th.eye()生成一个二维对角矩阵
        # 得到一个矩阵action_mask,用来将(episode_num, n_agents, n_agents * n_actions)的actions中每个agent自己的动作变成0向量
        action_mask = action_mask.view(-1, 1).repeat(1, self.n_actions).view(
            self.n_agents, -1)
        inputs.append(u_onehot * action_mask.unsqueeze(0))
        inputs_next.append(u_onehot_next * action_mask.unsqueeze(0))

        # 添加agent编号对应的one-hot向量
        '''
        因为当前的inputs三维的数据,每一维分别代表(episode编号,agent编号,inputs维度),直接在后面添加对应的向量
        即可,比如给agent_0后面加(1, 0, 0, 0, 0),表示5个agent中的0号。而agent_0的数据正好在第0行,那么需要加的
        agent编号恰好就是一个单位矩阵,即对角线为1,其余为0
        '''
        inputs.append(
            torch.eye(self.n_agents).unsqueeze(0).expand(episode_num, -1, -1))
        inputs_next.append(
            torch.eye(self.n_agents).unsqueeze(0).expand(episode_num, -1, -1))

        # 要把inputs中的5项输入拼起来,并且要把其维度从(episode_num, n_agents, inputs)三维转换成(episode_num * n_agents, inputs)二维
        inputs = torch.cat(
            [x.reshape(episode_num * self.n_agents, -1) for x in inputs],
            dim=1)
        inputs_next = torch.cat(
            [x.reshape(episode_num * self.n_agents, -1) for x in inputs_next],
            dim=1)
        return inputs, inputs_next

    def _get_q_values(self, batch, max_episode_len):
        episode_num = batch['o'].shape[0]
        q_evals, q_targets = [], []
        for transition_idx in range(max_episode_len):
            inputs, inputs_next = self._get_critic_inputs(
                batch, transition_idx, max_episode_len)
            if self.args.cuda:
                inputs = inputs.cuda()
                inputs_next = inputs_next.cuda()
            # 神经网络输入的是(episode_num * n_agents, inputs)二维数据,得到的是(episode_num * n_agents, n_actions)二维数据
            q_eval = self.eval_critic(inputs)
            q_target = self.target_critic(inputs_next)

            # 把q值的维度重新变回(episode_num, n_agents, n_actions)
            q_eval = q_eval.view(episode_num, self.n_agents, -1)
            q_target = q_target.view(episode_num, self.n_agents, -1)
            q_evals.append(q_eval)
            q_targets.append(q_target)
        # 得的q_evals和q_targets是一个列表,列表里装着max_episode_len个数组,数组的的维度是(episode个数, n_agents,n_actions)
        # 把该列表转化成(episode个数, max_episode_len, n_agents,n_actions)的数组
        q_evals = torch.stack(q_evals, dim=1)
        q_targets = torch.stack(q_targets, dim=1)
        return q_evals, q_targets

    def _get_actor_inputs(self, batch, transition_idx):
        # 取出所有episode上该transition_idx的经验,u_onehot要取出所有,因为要用到上一条
        obs, u_onehot = batch['o'][:, transition_idx], batch['u_onehot'][:]
        episode_num = obs.shape[0]
        inputs = []
        inputs.append(obs)
        # 给inputs添加上一个动作、agent编号

        if self.args.last_action:
            if transition_idx == 0:  # 如果是第一条经验,就让前一个动作为0向量
                inputs.append(torch.zeros_like(u_onehot[:, transition_idx]))
            else:
                inputs.append(u_onehot[:, transition_idx - 1])
        if self.args.reuse_network:
            # 因为当前的inputs三维的数据,每一维分别代表(episode编号,agent编号,inputs维度),直接在dim_1上添加对应的向量
            # 即可,比如给agent_0后面加(1, 0, 0, 0, 0),表示5个agent中的0号。而agent_0的数据正好在第0行,那么需要加的
            # agent编号恰好就是一个单位矩阵,即对角线为1,其余为0
            inputs.append(
                torch.eye(self.args.n_agents).unsqueeze(0).expand(
                    episode_num, -1, -1))
        # 要把inputs中的三个拼起来,并且要把episode_num个episode、self.args.n_agents个agent的数据拼成40条(40,96)的数据,
        # 因为这里所有agent共享一个神经网络,每条数据中带上了自己的编号,所以还是自己的数据
        inputs = torch.cat(
            [x.reshape(episode_num * self.args.n_agents, -1) for x in inputs],
            dim=1)
        # TODO 检查inputs_next是不是相当于inputs向后移动一条
        return inputs

    def _get_action_prob(self, batch, max_episode_len, epsilon):
        episode_num = batch['o'].shape[0]
        avail_actions = batch[
            'avail_u']  # coma不用target_actor,所以不需要最后一个obs的下一个可执行动作
        action_prob = []
        for transition_idx in range(max_episode_len):
            inputs = self._get_actor_inputs(
                batch, transition_idx)  # 给obs加last_action、agent_id
            if self.args.cuda:
                inputs = inputs.cuda()
                self.eval_hidden = self.eval_hidden.cuda()
            outputs, self.eval_hidden = self.eval_rnn(
                inputs, self.eval_hidden
            )  # inputs维度为(40,96),得到的q_eval维度为(40,n_actions)
            # 把q_eval维度重新变回(8, 5,n_actions)
            outputs = outputs.view(episode_num, self.n_agents, -1)
            prob = torch.nn.functional.softmax(outputs, dim=-1)
            action_prob.append(prob)
        # 得的action_prob是一个列表,列表里装着max_episode_len个数组,数组的的维度是(episode个数, n_agents,n_actions)
        # 把该列表转化成(episode个数, max_episode_len, n_agents,n_actions)的数组
        action_prob = torch.stack(action_prob, dim=1).cpu()

        action_num = avail_actions.sum(dim=-1, keepdim=True).float().repeat(
            1, 1, 1, avail_actions.shape[-1])  # 可以选择的动作的个数
        action_prob = ((1 - epsilon) * action_prob +
                       torch.ones_like(action_prob) * epsilon / action_num)
        action_prob[avail_actions == 0] = 0.0  # 不能执行的动作概率为0

        # 因为上面把不能执行的动作概率置为0,所以概率和不为1了,这里要重新正则化一下。执行过程中Categorical会自己正则化。
        action_prob = action_prob / action_prob.sum(dim=-1, keepdim=True)
        # 因为有许多经验是填充的,它们的avail_actions都填充的是0,所以该经验上所有动作的概率都为0,在正则化的时候会得到nan。
        # 因此需要再一次将该经验对应的概率置为0
        action_prob[avail_actions == 0] = 0.0
        if self.args.cuda:
            action_prob = action_prob.cuda()
        return action_prob

    def init_hidden(self, episode_num):
        # 为每个episode中的每个agent都初始化一个eval_hidden
        self.eval_hidden = self.eval_rnn.init_hidden().unsqueeze(0).expand(
            episode_num, self.n_agents, -1)

    def _train_critic(self, batch, max_episode_len, train_step):
        u, r, avail_u, terminated = batch['u'], batch['r'], batch[
            'avail_u'], batch['terminated']
        u_next = u[:, 1:]
        padded_u_next = torch.zeros(*u[:, -1].shape,
                                    dtype=torch.long).unsqueeze(1)
        u_next = torch.cat((u_next, padded_u_next), dim=1)
        mask = (1 - batch["padded"].float()).repeat(
            1, 1, self.n_agents)  # 用来把那些填充的经验的TD-error置0,从而不让它们影响到学习
        if self.args.cuda:
            u = u.cuda()
            u_next = u_next.cuda()
            mask = mask.cuda()
        # 得到每个agent对应的Q值,维度为(episode个数, max_episode_len, n_agents,n_actions)
        # q_next_target为下一个状态-动作对应的target网络输出的Q值,没有包括reward
        q_evals, q_next_target = self._get_q_values(batch, max_episode_len)
        q_values = q_evals.clone()  # 在函数的最后返回,用来计算advantage从而更新actor
        # 取每个agent动作对应的Q值,并且把最后不需要的一维去掉,因为最后一维只有一个值了

        q_evals = torch.gather(q_evals, dim=3, index=u).squeeze(3)
        q_next_target = torch.gather(q_next_target, dim=3,
                                     index=u_next).squeeze(3)
        targets = td_lambda_target(batch, max_episode_len, q_next_target.cpu(),
                                   self.args)
        if self.args.cuda:
            targets = targets.cuda()
        td_error = targets.detach() - q_evals
        masked_td_error = mask * td_error  # 抹掉填充的经验的td_error

        # 不能直接用mean,因为还有许多经验是没用的,所以要求和再比真实的经验数,才是真正的均值
        loss = (masked_td_error**2).sum() / mask.sum()
        # print('Loss is ', loss)
        self.critic_optimizer.zero_grad()
        loss.backward()
        torch.nn.utils.clip_grad_norm_(self.critic_parameters,
                                       self.args.grad_norm_clip)
        self.critic_optimizer.step()
        if train_step > 0 and train_step % self.args.target_update_cycle == 0:
            self.target_critic.load_state_dict(self.eval_critic.state_dict())
        return q_values

    def save_model(self, train_step):
        num = str(train_step // self.args.save_cycle)
        if not os.path.exists(self.model_dir):
            os.makedirs(self.model_dir)
        torch.save(self.eval_critic.state_dict(),
                   self.model_dir + '/' + num + '_critic_params.pkl')
        torch.save(self.eval_rnn.state_dict(),
                   self.model_dir + '/' + num + '_rnn_params.pkl')
Esempio n. 2
0
class COMA:
    def __init__(self, args):
        self.n_agents = args.n_agents
        self.n_actions = args.n_actions
        self.obs_shape = args.obs_shape
        actor_input_shape = self.obs_shape
        # Actor Network (RNN)
        if args.last_action:
            actor_input_shape += self.n_actions
        if args.reuse_networks:
            actor_input_shape += self.n_agents

        self.eval_rnn = RNN(actor_input_shape, args)
        print('Init Algo Coma')

        self.eval_critic = ComaCritic(args)
        self.target_critic = ComaCritic(args)

        if args.use_cuda and torch.cuda.is_available():
            self.device = torch.device("cuda:0")
            self.eval_rnn.to(self.device)
            self.eval_critic.to(self.device)
            self.target_critic.to(self.device)
        else:
            self.device = torch.device("cpu")

        self.target_critic.load_state_dict(self.eval_critic.state_dict())

        self.rnn_parameters = list(self.eval_rnn.parameters())
        self.critic_parameters = list(self.eval_critic.parameters())

        self.critic_optimizer = torch.optim.Adam(self.critic_parameters,
                                                 lr=args.critic_lr)
        self.rnn_optimizer = torch.optim.Adam(self.rnn_parameters,
                                              lr=args.actor_lr)

        self.args = args
        self.loss_func = torch.nn.MSELoss()
        self.eval_hidden = None

    def learn(self, batch, max_episode_len, train_step, epsilon):
        bs = batch['o'].shape[0]
        self.init_hidden(bs)
        for key in batch.keys():
            if key == 'u':
                batch[key] = torch.tensor(batch[key], dtype=torch.long)
            else:
                batch[key] = torch.tensor(batch[key], dtype=torch.float32)
        u, r, terminated = batch['u'], batch['r'], batch['terminated']
        if self.args.use_cuda:
            u = u.to(self.device)

        critic_rets = self._train_critic(batch, train_step)
        q_taken, q_values = [], []
        for a_i, (q_eval, q_all) in zip(range(self.n_agents), critic_rets):
            q_taken.append(q_eval)
            q_values.append(q_all)
        q_taken = torch.stack(q_taken, dim=2).squeeze(3)
        q_values = torch.stack(q_values, dim=2)
        action_prob = self._get_action_prob(batch, max_episode_len, epsilon)

        pi_taken = torch.gather(action_prob, dim=3, index=u).squeeze(3)
        log_pi_taken = torch.log(pi_taken)

        # Advantage for actor(policy) optimization
        baseline = (q_values * action_prob).sum(
            dim=3, keepdim=True).squeeze(3).detach()
        advantage = (q_taken - baseline).detach()
        loss = -(advantage * log_pi_taken).mean()
        self.rnn_optimizer.zero_grad()
        disable_gradients(self.eval_critic)
        loss.backward()
        enable_gradients(self.eval_critic)
        torch.nn.utils.clip_grad_norm_(self.rnn_parameters,
                                       self.args.grad_norm_clip)
        self.rnn_optimizer.step()

    def _train_critic(self, batch, train_step):
        """
        Unlike the qmix or vdn which seems like q_learning to choose the argmax Q values as the q_targets
        COMA is someway like the MADDPG or DDPG algorithm which is deterministic policy gradient method
        So it requires the deterministic next action infos as 'u_next'

        :return: [n_agents * [(bs, episode_limit, 1), (bs, episode_limit, n_actions)]]
        """
        r, terminated = batch['r'], batch['terminated']
        if self.args.use_cuda:
            r = r.to(self.device)
            terminated = terminated.to(self.device)

        critic_in, target_critic_in = self._get_critic_inputs(batch)
        q_targets = self.target_critic(
            target_critic_in)  # n_agents * (bs, episode_limit, 1)
        critic_rets = self.eval_critic(critic_in, return_all_q=True)
        q_loss = 0
        for a_i, q_target, (q_eval, q_all) in zip(range(self.n_agents),
                                                  q_targets, critic_rets):
            target = r + self.args.gamma * q_target * (1 - terminated)
            q_loss += self.loss_func(target, q_eval)

        self.critic_optimizer.zero_grad()
        q_loss.backward()
        self.eval_critic.scale_shared_grads()
        torch.nn.utils.clip_grad_norm_(
            self.eval_critic.parameters(),
            self.args.grad_norm_clip * self.n_agents)
        self.critic_optimizer.step()

        if train_step > 0 and train_step % self.args.target_update_cycle == 0:
            self.target_critic.load_state_dict(self.eval_critic.state_dict())

        return critic_rets

    def _get_critic_inputs(self, batch):
        """
        The COMA algorithm handle the critic inputs with total steps (without transition_idx)
        """
        obs, obs_next = batch['o'], batch[
            'o_next']  # (bs, episode_limit, n_agents, obs_shape)
        u_onehot = batch[
            'u_onehot']  # (bs, episode_limit, n_agents, n_actions)
        u_onehot_next = u_onehot[:,
                                 1:]  # (bs, episode_limit - 1, n_agents, n_actions)
        padded_next = torch.zeros(*u_onehot[:, -1].shape,
                                  dtype=torch.float32).unsqueeze(
                                      1)  # Add a step with zeros
        u_onehot_next = torch.cat((u_onehot_next, padded_next), dim=1)
        if self.args.use_cuda:
            obs = obs.to(self.device)
            obs_next = obs_next.to(self.device)
            u_onehot = u_onehot.to(self.device)
            u_onehot_next = u_onehot_next.to(self.device)

        agents_obs, agents_obs_next = [], []
        agents_u, agents_u_next = [], []
        for a_i in range(self.n_agents):
            agent_obs, agent_obs_next = obs[:, :,
                                            a_i], obs_next[:, :,
                                                           a_i]  # (bs, episode_limit, obs_shape)
            agent_u, agent_u_next = u_onehot[:, :,
                                             a_i], u_onehot_next[:, :,
                                                                 a_i]  # (bs, episode_limit, n_actions)
            agents_obs.append(agent_obs)
            agents_obs_next.append(agent_obs_next)
            agents_u.append(agent_u)
            agents_u_next.append(agent_u_next)

        target_critic_in = list(zip(agents_obs_next, agents_u_next))
        critic_in = list(zip(agents_obs, agents_u))

        return critic_in, target_critic_in

    def _get_actor_inputs(self, batch, transition_idx):
        # Because the rnn agent actor network didn't initialize a target network, it requires none next infos
        obs, u_onehot = batch['o'][:, transition_idx], batch['u_onehot'][:]
        bs = obs.shape[0]
        # Observation
        inputs = [obs]

        if self.args.last_action:
            if transition_idx == 0:
                inputs.append(torch.zeros_like(u_onehot[:, transition_idx]))
            else:
                inputs.append(u_onehot[:, transition_idx - 1])
        if self.args.reuse_networks:
            inputs.append(
                torch.eye(self.args.n_agents).unsqueeze(0).expand(bs, -1, -1))
        # Since the using of GRU network, the inputs shape should be shaped as 2 dimensions
        inputs = torch.cat(
            [x.reshape(bs * self.args.n_agents, -1) for x in inputs], dim=1)
        return inputs

    def _get_action_prob(self, batch, max_episode_len, epsilon):
        bs = batch['o'].shape[0]
        # The available actions for each agent. In MPE, an agent could choose every action at any time-step.
        avail_actions = torch.ones_like(
            batch['u_onehot']
        )  # (bs, episode_limit, n_agents, n_actions) --> all 1
        action_prob = []
        for transition_idx in range(max_episode_len):
            inputs = self._get_actor_inputs(batch, transition_idx)
            if self.args.use_cuda:
                inputs = inputs.to(self.device)
                self.eval_hidden = self.eval_hidden.to(self.device)
            outputs, self.eval_hidden = self.eval_rnn(
                inputs, self.eval_hidden)  # outputs:(bs * n_agents, n_actions)
            outputs = outputs.view(bs, self.n_agents, -1)
            prob = torch.nn.functional.softmax(outputs, dim=-1)
            action_prob.append(prob)

        action_prob = torch.stack(
            action_prob,
            dim=1).cpu()  # (bs, episode_limit, n_agents, n_actions)
        actions_num = avail_actions.sum(dim=-1, keepdim=True).float().repeat(
            1, 1, 1, avail_actions.shape[-1])
        action_prob = (1 - epsilon) * action_prob + torch.ones_like(
            action_prob) * epsilon / actions_num
        action_prob = action_prob / action_prob.sum(dim=-1, keepdim=True)
        if self.args.use_cuda:
            action_prob = action_prob.to(self.device)
        return action_prob

    def _td_lambda_target(self, batch, max_episode_len, q_targets):
        bs = batch['o'].shape[0]
        terminated = (1 - batch["terminated"].float()).repeat(
            1, 1, self.n_agents)
        r = batch['r'].repeat(
            (1, 1, self.n_agents
             ))  # (bs, episode_limit, 1) --> (bs, episode_limit, n_agents)

        n_step_returns = torch.zeros(
            (bs, max_episode_len, self.n_agents, max_episode_len))
        for transition_idx in range(max_episode_len - 1, -1, -1):
            n_step_returns[:, transition_idx, :, 0] = r[:, transition_idx] + self.args.gamma *\
                                                      q_targets[:, transition_idx] * terminated[:, transition_idx]
            for n in range(1, max_episode_len - transition_idx):
                n_step_returns[:, transition_idx, :, n] = r[:, transition_idx] + self.args.gamma *\
                                                          n_step_returns[:, transition_idx + 1, :, n - 1]

        lambda_return = torch.zeros((bs, max_episode_len, self.n_agents))
        for transition_idx in range(max_episode_len):
            returns = torch.zeros((bs, self.n_agents))
            for n in range(1, max_episode_len - transition_idx):
                returns += pow(self.args.td_lambda, n -
                               1) * n_step_returns[:, transition_idx, :, n - 1]
                lambda_return[:, transition_idx] = (1 - self.args.td_lambda) * returns +\
                                                   pow(self.args.td_lambda, max_episode_len - transition_idx - 1) *\
                                                   n_step_returns[:, transition_idx, :, max_episode_len - transition_idx - 1]

        return lambda_return

    def init_hidden(self, batch_size):
        self.eval_hidden = torch.zeros(
            (batch_size, self.n_agents, self.args.rnn_hidden_dim))

    def get_params(self):
        return {
            'eval_critic': self.eval_critic.state_dict(),
            'eval_rnn': self.eval_rnn.state_dict()
        }

    def load_params(self, params_dict):
        # Get parameters from save_dict
        self.eval_rnn.load_state_dict(params_dict['eval_rnn'])
        self.eval_critic.load_state_dict(params_dict['eval_critic'])
        # Copy the eval networks to target networks
        self.target_critic.load_state_dict(self.target_critic.state_dict())