Skip to content

A new version of Deep Q-Learning designed for sparse reward environments and tested in a grid world.

Notifications You must be signed in to change notification settings

Toroi01/DQL-B-C-FrozenLake

Repository files navigation

DQL-B-C-FrozenLake

State-of-the-art methods for deep reinforcement learning have demonstrated the ability to learn complex game strategies for Atari’s games, Chess, Go... Deep Q-Learning was the first deep reinforcement learning method to outperform the human level at some of Atari’s games, however, it performs poorly in environments with sparse rewards or with the facility to die. In both cases, the algorithm is not able to learn from low probability rewards. In this project, a modified version of Deep-Q Learning which implements Deep Q-Learning with Boltzmann Count-Based exploration (DQL-B-C) has been developed. The algorithm has been tested and compared in a grid environment with different grid sizes. The experiments demonstrate that DQL-B-C explores better and learns more quickly.

Slides: https://docs.google.com/presentation/d/1aFe31BxnpFCj2PPX7y-xMYof2UImC5n40bP7NmLgLp8/edit?usp=sharing

Demo: https://youtu.be/agRfDT3sp9A

About

A new version of Deep Q-Learning designed for sparse reward environments and tested in a grid world.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published