Robert Carter Undergraduate Dissertation 2017/18

Building a Deep Learning Agent for Battleship

Supervised by E.Vasilaki

Abstract

Reinforcement learning is an area of machine learning which is focused on how software agents take actions within an environment to maximize a reward function, the aim of this is for the agent to determine the ideal behaviour within its environment, to maximize its performance. In 2015 Google deep mind developed a reinforcement learning agent capable of learning and playing Atari 2600 classic games using high dimensional sensory inputs. This agent surpassed skilled human ability within many of the games tested.

This project aims to develop a game of battleships then apply several deep reinforcement learning tachniques in order to create an agent that can play the game and learn through play to improve with an aim of outperforming current methods for agents to play the game.