Transfer Learning & Neuro-Evolution for Atari Games

Introduction

Agents in reinforcement learning tasks need to transform high dimensional sensory-inputs from the environment into an action to take. This can be performed using neural networks. Neuro-evolution is the act of improving the neural network by a genetic algorithm rather than gradient-based learning which requires a large labelled dataset. Transfer learning is often used by convolutional neural networks for computer vision, allowing the basic features learnt by the models to be transferred over to other tasks.

Method

This project aims to test the transfer learning capabilities of deep neural networks in the popular reinforcement learning tasks of learning to play Atari games. In particular, the models will be trained on Space Invaders, and the learning will be transferred to Demon Attack and Carnival, two games with a similar goal and require the same actions being taken (left, right, shoot).

Key Technologies

• Tensorflow
• Python

Links

• GitHub