nicofirst1 / MAMBRL

Multi agent model based reinforcment learning

Installation

This project requires python>= 3.6

First clone the repo with"

git clone --recurse-submodules https://github.com/nicofirst1/MAMBRL.git

We are using a custom version of the PettingZoo library, in order to install it execute:

pip install PettingZoo/
pip install .

If you wish to modify the library while coding use pip install -e . instead.

Training on colab

You can train the model on

The notebook uses weight and biases in order to keep track of the experiment.

Problem Statement

Description of the Problem:

Abstract Definition:

Consider 'N' points on a 2D plane which need to be visited at least once by 'M' bots. Provide a path for all of the 'M' bots such that the total distance travelled by the bots is minimum.

Consider the following

Target Points P1, a set of ‘N’ points P1 = { p1, p2, p3, … pN }

Bots P2, a set of 'M' points P2 = { b1, b2, b3, ... bM }

M<<N

Return a matrix Y which gives the sequence of points needed by each of the bots to cover all the points minimizzing the total distace travelled.

Y[i,j] = ith point in the path of bot j.

Attempt to Solution

This project aims to design and test a customized model of Markov Decision Process and attempt to solve it using Cooperative Q-Learning.

Reasons for Multi Agent Reinforcement Learning to be the best approach for solution

M<<N

1. Consider the use case where M=50 (50 bots) and N=1000 (1000 Target Points)
2. The total number of possible paths is in order of 10^90 which is much greater than the generic use cases of 20 possible paths
3. The environment is unexplored, meaning that location of all the 1000 points is unknown, The bots need to explore and simultaneously learn to move efficiently in the environment. Absence of exact knowledge of environment rules out the validity of deterministic algorithms.

In this project, the notion of multi-agent envronment is implemented using a novel method which we call ConnectedQ

Model based RL

• Check this out for an overview on model based RL. Specifically we refer to the data generation technique.

TODOs

• [] add tune configurations
• [] dockerize