DQNs

Implementation of DQNs.

Environment : OpenAI Gym Atari 2600 games

Papers

DQN : Playing Atari with Deep Reinforcement Learning

Double DQN : Deep Reinforcement Learning with Double Q-learning

Prioritized Replay : PRIORITIZED EXPERIENCE REPLAY

Dueling Network : Dueling Network Architectures for Deep Reinforcement Learning

Ape-X DQN : DISTRIBUTED PRIORITIZED EXPERIENCE REPLAY

Usage

$ python dqn_atari.py --prioritized 1 --double 1 --dueling 1 --n_step 3

Prioritized Experience Replay

--prioritezed : 0 or 1

Double Deep Q Learning (DDQN)

--double : 0 or 1

Dueling Network

--dueling : 0 or 1

multi-step bootstrap target

--n_step : int (1 : normal TD error)

Other arguments are described in dqn_atari.py

Ape-X DQN

See https://github.com/omurammm/apex_dqn

Results

After 12,000 episodes (Ape-X DQN)

Learning curves