LLM-Based Multi Agent System and framework to play game theoretic games

This architecture is a modular framework to test different agents playing game theoretic games in different settings. The implementation proposes the Prisoner's Dilemma against deterministic opponents. This project has been developed by Gabriele Nanni and Davide Cremonini.

Available Players

The available Player architectures are:

Plain LLM: The agents consists in a single LLM agent. Tool supported LLM: The LLM agent is supported with tools to determine the best game theoretical choice. LLM-based MAS: Two LLMs (Decision Maker and Assistant) discuss about the best strategy to deploy. <\itemize>

The available Opponent strategies are:

Always Cooperate: the opponent cooperates blindly. Always Defect: the opponent defects blindly Tit For Tat: the opponent starts with cooperation and then plays whatever the player chose in the previous round. <\itemize>

The possible settings are:

Classic prisoner's Dilemma. Business setting. Strategic game setting. <\itemize>

The Experiment runs different payoff configurations, all the parameters reguarding player architecture, opponent strategy and setting are to be set in the config file.