Subtract the entropy to encourage exploration.

e9dbef6b · Mark Daoust · GitHub · 949b1987 · e9dbef6b
Unverified Commit e9dbef6b authored Aug 30, 2018 by Mark Daoust Committed by GitHub Aug 30, 2018
Show whitespace changes
Inline Side-by-side

Showing with 1 addition and 1 deletion

research/a3c_blogpost/a3c_cartpole.py research/a3c_blogpost/a3c_cartpole.py +1 -1

No files found.
--- a/research/a3c_blogpost/a3c_cartpole.py
+++ b/research/a3c_blogpost/a3c_cartpole.py
@@ -353,7 +353,7 @@ class Worker(threading.Thread):
    policy_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=memory.actions,
                                                                 logits=logits)
    policy_loss *= tf.stop_gradient(advantage)
-    policy_loss = 0.01 * entropy
+    policy_loss -= 0.01 * entropy
    total_loss = tf.reduce_mean((0.5 * value_loss + policy_loss))
    return total_loss