Attention-Actor-Critic-for-Atari-Learning

A Reinforcement Learning algorithm with attention mechanism for Atari Learning

This is a Reinforcement Learning based Actor Critic model for Atari Games Learning. The repository contains three different architectures.

Attention Model

Uses dot product attention mechanism as explained in the paper "Attention is all you need" in an Actor Critic network to predict future actions. Requires one frame as state input

Attention Visualization

Recurrent Model

Uses a Gated Recurrent Unit to understand temporal dependencies in order to predict future actions. Requires one frame as state input.

Convolutional Model

conv-a3c.py uses convolutional network to predict optimal actions. It uses four frames as state input.