RL: Tricks of the TradeGetting RL to work is hard. Collecting some tricks here.
Jul 5, 2020Last updated: Jul 5, 2020
Raw collection of random things I've read around for better RL but now forgotten the sources. Hopefully, I can refine this sometime.
- Off-policy learning
- Flexible function approx.
- Experience replay buffer + mini-batch SGD
- Target network
- TD-error clipping
- Double Q-Learning - reduce maximization bias
- Average Q-Learning - reduce variance
Optimistic initializations - initialize to upper bound of Q-values