Reinforcement Learning in Pacman

Introduction

In this project experimented with various MDP and Reinforcement Learning techniques namely value iteration, Q-learning and approximate Q-learning. This is part of Pacman projects developed at UC Berkeley.

Directory Structure

---RL

       qlearningAgents.py

       analysis.py

---lab.pdf

---README.md

---report.pdf

Executing

Then run the autograder using $python autograder.py

It gave me a score of 25/25.

Value Iteration

$python gridworld.py -a value -i 100 -k 10

$python gridworld.py -a value -i 5

Bridge Crossing Analysis

$python gridworld.py -a value -i 100 -g BridgeGrid --discount 0.9 --noise 0.2

Policies

$python autograder.py -q q3

Q-Learning

python gridworld.py -a q -k 5 -m

Epsilon Greedy

$python gridworld.py -a q -k 100

$python crawler.py

Bridge Crossing Revisited

$python gridworld.py -a q -k 50 -n 0 -g BridgeGrid -e 1

Q-Learning and Pacman

$python pacman.py -p PacmanQAgent -x 2000 -n 2010 -l smallGrid

Approximate Q-Learning

$python pacman.py -p ApproximateQAgent -x 2000 -n 2010 -l smallGrid

$python pacman.py -p ApproximateQAgent -a extractor=SimpleExtractor -x 50 -n 60 -l mediumGrid

$python pacman.py -p ApproximateQAgent -a extractor=SimpleExtractor -x 50 -n 60 -l mediumClassic

Developed by

Sai Srinadhu K