README.md

keras-only-RL

Implementating Reinforcement Learning from A to Z using keras only.

How to run

For example: Run Atari Breakout with D3QN Agent

sh atari_breakout_run.sh --double=True --dueling=True

Help

sh atari_breakout_run.sh -h

usage: main.py [-h] [--e E] [--double D] [--dueling B]

Some hyperparameters

optional arguments:
  -h, --help   show this help message and exit
  --e E        Total episodes
  --double D   Enable Double DQN
  --dueling B  Enable Dueling DQN

To the Rainbow

Technique	Problem	How to solve it
DQN	Non-stationary targets makes learning unstable	Fixed Q-targets
	Correlation between samples makes W biased	Replay Memory
Double ~	Maximum estimator raises over-estimation	Using double estimators
Dueling ~	Some state may have inherently low value	Q(s, a) = V(s) + A(s, a)