Ephemeral Value Adjustments based on pfrl

• from https://arxiv.org/abs/1810.08163

I implemented an algorithm called Ephemeral Value Adjustments (EVA) using pfrl.

Japanese README

EVA

• Fast deep reinforcement learning using online adjustments from the past

Requirement

• python >= 3.8
  • pfrl
  • OpenAI gym
    • atari-py >= 0.2.6

Usage

python train.py [options]

Options

• --env : Environment name of the Atari.
  • Eg : --env PongNoFrameskip-v4
• --gpu : If 0, use the GPU. If the number is less than 0, the GPU is not used.
• --agent : Choose DQN and EVA.

TODO

• Reduce GPU memory usage
• More experiments.

Performance Comparison