Reinforcement Learning algorithms in a Mean field games with a lot of players

In this thesis we look at Reinforcement Learning (RL) algorithms improvements for Mean Field Games. Mean field game (MFG) theory is the study of the stochastic decision making in a game with large number of agents where the interactions between agents are negligible, but each agent's action affects general environment state. Reinforcement learning methods in general solves the easy version of the game, where the main assumption is that all agents do not interact to each other which is generally not true. The goal of the thesis is to find the modification of Reinforcement learning algorithm that will consider not every agent individually but more accurate game definition. The main idea for improvement is to consider a multiagent game as a mean field game and consider one arbitrary agent as representative agent and all other agents as a population state-action distribution.