This repository contains the code pertaining to the Interactive Multi-Agent Reinforcement Learning project done under Computer Society IEEE NITK.

Multi-Agent Reinforcement Learning (MARL) deals with multiple cooperative, competitive, or mixed agents interacting to achieve goals in a shared environment. Extension of single-agent optimization workhorses like Gradient Descent Ascent (GDA) is one straightforward way for multi-agent optimization. However, this ignores non-stationarity and performs poorly. Hence in this project, we studied and implemented Competitive Gradient Descent (CGD) and Learning with Opponent Learning Awareness (LOLA) algorithms that solve this problem in multi-agent environments. We implemented these algorithms in a wide variety of multi-agent settings and social dilemmas like the bilinear game, linear-quadratic control, matching pennies, rock paper scissors, prisoner’s dilemma, iterated prisoner’s dilemma, Markov soccer, and the coin game.