Achieving collaborative behaviors through multi-agent reinforcement learning in Unity

Abstract

Programming human-like cooperative behaviors in enemies or non-player characters in videogames has been a highly pursued objective over the years. Traditionally, such behaviors have been implemented following complex algorithms and heuristics requiring extensive prior expertise and highly dependent on the game mechanics and its casuistry. Recently, however, machine learning techniques such as reinforcement learning (from now on RL) have opened the door to the implementation of relatively strong AIs without having to implement and hand-tune complex dedicated algorithms. This project tries to cover the case of collaborative behaviors using machine learning, which has been relatively less explored nowadays. The main objective of this project is to train different agents using Machine Learning to be able to compete in a simple shooter game with two teams of two players. Using multi-agent RL, each agent will practice against copies of himself (self-play) and, through multiple generations, it will learn how to interact with the environment and their own teammates to cooperate in order to defeat the enemy team. The base of this project will be built in Unity and, to handle Machine Learning, their own ML Agents tool will be used. The aim is to compare different strategies towards implementing collaborative behavior and to create a front-end in Unity capable of guiding the user to the creation of collaborative AIs.