alpha_w = 1e-3  # 初始化

    optimizer_w = optim.Adam(**s_value_func**.parameters(), lr=alpha_w)
    optimizer_w.zero_grad()
    policy_loss_w =-delta
    policy_loss_w.backward(retain_graph = True)
    clip_grad_norm_(policy_loss_w, 0.1)
    optimizer_w.step()