关于代码问题的描述

[Camellia77]

def process_env_step(self, rewards, costs, dones, infos):

    self.transition.rewards = rewards.clone()
    self.transition.costs = costs.clone()
    self.transition.dones = dones
    # Bootstrapping on time outs
    if 'time_outs' in infos:
        self.transition.rewards += self.gamma * torch.squeeze(self.transition.values * infos['time_outs'].unsqueeze(1).to(self.device), 1)
        self.transition.costs += self.gamma * (self.transition.costs * infos['time_outs'].unsqueeze(1).to(self.device))
    # Record the transition
    self.storage.add_transitions(self.transition)
    self.transition.clear()
    self.actor_critic.reset(dones)  首先很感谢您的开源，有点关于np30算法中的问题想询问一下，就是这一段关于time_out的损失积累中，self.transition.costs += self.gamma * (self.transition.costs * infos['time_outs'].unsqueeze(1).to(self.device))，为什么不是仿照reward一样用cost_values积累，而是采用costs呢，这样做不是相当于在time_out的时候手动对costs进行放大吗（costs = (1+gamma)*costs ),这个地方不是很理解 希望能得到解答