search

muupan / dqn-in-the-caffe

An implementation of Deep Q-Network using Caffe
MIT License
213 stars 118 forks source link

Parameters used to learn Pong? #1

Closed mhauskn closed 9 years ago

mhauskn commented 9 years ago

What parameters were used to learn the pong player showed in the video (https://www.youtube.com/watch?v=p88R2_3yWPA)? Specifically what is the gamma/memory size/iterations needed to train dqn? I attempted to retrain the agent using the default parameters but it shows no progress after 2-million iterations.

Thanks in advance, and very cool repo!

muupan commented 9 years ago

Thanks for forking my repo! Default parameters in dqn_solver.prototxt was outdated, so I've just updated it and now it's the one used in the Pong demo. The size of replay memory was 500,000.

mhauskn commented 9 years ago

Thanks, it seems to be working much better now :)