Learning to Deceive in Multi-Agent Hidden Role Games

TL;DR

This paper introduces a complex multi-agent reinforcement learning environment inspired by social deception games, along with a Bayesian belief manipulation model that enhances deception and agent performance.

Contribution

It presents a novel multi-agent environment for studying deception and introduces a Bayesian belief manipulation model to improve deceptive strategies.

Findings

BBM effectively deceives other agents in the environment

Deceptive agents with BBM outperform non-deceptive agents

The environment models real-world social deception scenarios

Abstract

Deception is prevalent in human social settings. However, studies into the effect of deception on reinforcement learning algorithms have been limited to simplistic settings, restricting their applicability to complex real-world problems. This paper addresses this by introducing a new mixed competitive-cooperative multi-agent reinforcement learning (MARL) environment inspired by popular role-based deception games such as Werewolf, Avalon, and Among Us. The environment's unique challenge lies in the necessity to cooperate with other agents despite not knowing if they are friend or foe. Furthermore, we introduce a model of deception, which we call Bayesian belief manipulation (BBM) and demonstrate its effectiveness at deceiving other agents in this environment while also increasing the deceiving agent's performance.