Closed kvfrans closed 8 years ago
Nice work. Few comments:
eligibility = tf.log(good_probabilities)
" -- it's not really the eligibility until you multiply it by the advantage. I recommend calling it something else.advantages = tf.placeholder("float",[None,1])
eligibility = tf.log(good_probabilities) * advantages
is hard to connect to the previous code. In particular, it's not obvious that you differentiate eligibility, at least not form a shallow reading. The second point is more important, and makes it possible to change the name of the first eligibility. Basically, don't shy away form a bit of code duplication in the explanation to your solution. Otherwise, it's good.
Let me know when it's done and I'll merge it.
Thanks for the feedback, I updated the post.
Writeup and code implementing random search, hill climbing, and policy gradient on the cartpole environment