This project was worked on as part of Stanford’s CS 238: Decision Making Under Uncertainty, led by Professor Mykel Kochenderfer. Abstract, full report, and code repo available below.

Abstract

Modelling sports can be especially challenging given the large variety and high complexity of factors that human professional athletes take into account simultaneously when making decisions during gameplay. In this paper, we first develop a parametrized simulation of the game of Tennis, using all modern rules of the game. Subsequently, we employ Reinforcement Learning to train sophisticated agents and evaluate their success in learning obvious as well as empirical rules about how to optimize their play, given their different capabilities. Using a null action-selection model as a baseline, we compare the performance of agents trained with Q-Learning to see how they perform against equally skilled players, as well as how they might adapt to exploit their personal strengths and mitigate their weaknesses.

Project Report: Final Project Report

Github link: Modelling Optimal Tennis Decision-Making with Reinforcement Learning