From 5910e08672173aaa7e68109467d3d93b4087d251 Mon Sep 17 00:00:00 2001 From: haoshengzou Date: Thu, 25 Jan 2018 10:11:36 +0800 Subject: [PATCH] data/utils.py added but not pushed... --- tianshou/core/opt.py | 1 + 1 file changed, 1 insertion(+) diff --git a/tianshou/core/opt.py b/tianshou/core/opt.py index 59e4608..51d9e1c 100644 --- a/tianshou/core/opt.py +++ b/tianshou/core/opt.py @@ -15,6 +15,7 @@ def DPG(policy, action_value): grad_ys = tf.gradients(critic_value_loss, critic_action_input) grad_policy_vars = tf.gradients(policy_action_output, trainable_variables, grad_ys=grad_ys) + # TODO: this is slightly different from ddpg implementations in baselines, keras-rl and rllab. it uses sampled action (with noise) rather than directly connect the two networks grads_and_vars = zip(grad_policy_vars, trainable_variables)