Project information

• Category: Unity
• Genre: Machine Learning
• Project date: April, 2022
• Project URL: GitHub Repository

A machine learning dissertation using ML Agents to determine the usefulness of machine learning in social deduction games.

This project showcases two scenarios of machine learning being used to form the foundations of a social deduction AI. The first scene showing an AI learning to pick out an "imposter" based on movement differences. The power to identify irregularities and build suspicion is essential for social deduction games, and this demo begins to prove that machine learning can be taught this with adjustable levels of accuracy.

The second scene showing cooperative behaviour between 3 different agents, all operating on the same brain. 2 agents must sacrifice themselves on buttons for the last to reach the end goal. Teamwork and cooperation is also a significant skill of social deduction games, and this demo shows that agents can work together to complete a common goal. Moreover, it shows that machine learning agents can reliably make decisions based on a changing environment. Despite all agents in this puzzle sharing the same brain, different agents will press different buttons, proving that each of them understand the goal, regardless of their individual role.

Whilst this project doesn't show a complete AI player, it does set foundations that can be built upon, suggesting the machine learning can be used as a tool for creating more fun and dynamic AI for games that typically don't use AI due to its complexity. All of which can be done using a popular game engine, to further support a more mainstream use for machine learning in games.

• Machine Learning for Game AI
• Unity ML-Agents