Imitating reactive human behaviours in games using neural networks

Treball final de Grau en Disseny i Desenvolupament de Videojocs. Codi: VJ1241. Curs acadèmic: 2019/2020Deep learning has allowed to create neural networks that can play any game almost optimally. However, not so many have been trained to play like humans, or more concretely, like one specific person. Most people have recognizable ways of playing specific games, and imitating those behaviors would allow to create bots that don’t appear to be artificially generated. Also, by imitating one person behaviors it would be easy to create bots that play at the same level of quality. This document, which is a Final Degree Work report for the Bachelor’s Degree in Video Game Design and Development, presents some techniques to create neural networks that can imitate human behaviors using Unity’s ML Agents SDK, an analysis on what behaviors can be modeled more precisely, what are the training costs and how good are the results