Ego Networks

Andrei Marin, Traian Rebedea, Ionel Hosu
Politehnica University of Bucharest

Abstract

Games on the Atari 2600 platform have served as a benchmark for reinforcement learning algorithms in recent years, and while deep reinforcement learning approaches make progress on most games, there are still some games that the majority of these algorithms struggle with. These are called hard exploration games. We introduce two new developments for the Random Network Distillation (RND) architecture. We apply self-attention and the mechanism of ego motion on the RND architecture and we evaluate them on three hard exploration tasks from the Atari platform. We find that the proposed ego network model improve the baseline of the RND architecture on these tasks.

Installation Guide

First install the conda environment

conda create --name <env_name> --file conda_requirements.txt

Then install dependencies that cannot be installed with conda

pip install -r pip_requirements.txt

Usage

To train an Ego RND agent on Montezuma's Revenge, run the following command

python run_atari.py --save_model

Acknowledgement

This work is based on OpenAI's Exploration by Random Network Distillation by Yuri Burda, Harri Edwards, Amos Storkey, Oleg Klimov