Measuring the impact of reinforcement learning on an electrooculography-only computer game

Abstract

In this paper we present an approach for detecting ocular movements, based on Electrooculographic (EOG) signals, that can have applications requiring the detection of ocular events such as saccades and blinks. We use it to implement an interactive go-kart game in which the user's goal is to avoid obstacles. Since horizontal saccades are the most representative of ocular movements, we use them as the main input for driving the kart. Eye blinking is a semi-autonomic and essential function that occurs naturally, so we decided to take advantage of it by using it as a secondary input to control the speed of the kart. This interface allows us to test the influence of machine learning techniques on game operation by inexperienced users and to evaluate whether it has a subjective positive impact. Two different versions of the game were implemented, one with a Reinforcement Learning Algorithm (RLA) that moderates users' commands based on outcomes of past commands, trying to prevent collisions, and a version with direct control (without RLA). Five participants tested the two versions of the game, so that we could compare the player's performance and engagement. We obtained promising results that show an improvement in score when RL is applied. We also found that players do not experience significant changes in gameplay feeling when RL is introduced.