runner.py 文件源码

def learn_single(self, value, value_last, last_action, reward):
        expected_value = self.gamma * value + reward # What value_last should have been if it was perfect

        value_loss = F.smooth_l1_loss(expected_value, value_last)
        print(value_loss.data)
        last_action.reinforce(value_loss.data[0])

        self.optimizer.zero_grad()
        final_nodes = [value_loss, last_action]
        gradients = [maybe_cuda(torch.ones(1)), None]
        autograd.backward(final_nodes, gradients, retain_graph=True)
        self.optimizer.step()
        del last_action