sentoa / rl-atari

Prioritized Experience Replay for A2C

some intro and upcoming link to paper

Files

Model	File	Description
DQN with Experience Replay	DQN.py	Deep Q-network algorithm from the Deepmind Paper
	DQN_PROD.py	Since DQN hyperparams requires a lot of memory, this use a downscaled replay memory size for production.
DQN with Priortized Experience Replay	DQNPER.py	From the Prioritzed Experience Replay Paper (need ref)
A2C	A2C.py	Actor-Critic Algorithm in vanilla form for Atari
A2C with Episodic Replay	A2CER.py	Actor-Critic Algorithm with Episodic Replay Memory
A2C with Prioritized Episodic Replay	A2CPER.py	Actor-Critic Algorithm with Prioritized Episodic Replay Memory

Requirements

TBD

rl-atari's Issues

Preserve framecount when resuming training

When the model saves, it should preserve frame and episode count, so when the model is loaded, it resumes where it last saved.

Save new model every 100 episode instead of overwriting

Add Arguments for training