Deriving MBTI Personalities in Multi-Agent Systems

Abstract

Research shows that diverse teams perform better than homogeneous ones. The differences in personalities of team members act as features of the team. For instance, in a predator- prey setting, an extroverted prey might operate in higher-density regions and eventually might get caught by the predators. An introverted prey though, would want to be isolated and as a result perform better in this domain. There have been multiple attempts to in- corporate diversity in personalities to multi-agent systems. (Salvit & Sklar 2012) use the Myers-Briggs model as a basis to encode personalities in agents in a NetLogo (Wilensky 1997b) Termites environment. Other attempts at personality modelling have been tightly coupled with the environment and agent design. In this thesis, we propose an environment- agnostic and agent-agnostic method of encoding the Myers-Briggs Type Indicator (MBTI) personalities in multi-agent systems by using techniques like reward shaping.We also show that diverse teams perform better than homogeneous ones since it is easier to control agent behaviours depending on whether the environment is a competitive or co-operative one. Keywords: Reinforcement Learning, Multi-Agent Systems, Myers-Briggs, Reward Shap- ing, Multi-Agent Deep Deterministic Policy Gradients