Nagy, Boldizsár (2024) Learning in zero-sum Stochastic Games. Bachelor's Thesis, Computing Science.
![[img]](https://fse.studenttheses.ub.rug.nl/style/images/fileicons/text.png) |
Text
bCS2024BoldizsarNagy.pdf Restricted to Registered users only Download (1MB) |
Abstract
The study of zero-sum stochastic games provides a framework to model and analyze competitive environments where strategic decision-making is crucial. The goal of modelling scenarios through the lens of stochastic games is to find the most rewarding action of the agents at any given state, which is derived from the optimal policy. This paper explores the application of reinforcement learning algorithms, specifically Value Iteration and Q-learning, within the context of stochastic games. Additionally, the project presents an engineering contribution as a proof of concept, an enhanced version of the Q-learning algorithm that utilizes Deep Reinforcement Learning techniques to solve these games effectively. While many Deep Q-learning solutions exist today for Markov Decision Processes (MDPs), these implementations are restricted to single agents interacting with the environment. However, this project focuses on extending these methods within an existing framework to zero-sum stochastic games which involve two players competing against each other. Based on the experiments in this project, the algorithm proved to be a successful proof of concept, demonstrating positive results on small benchmarks and affirming its potential for further development.
| Item Type: | Thesis (Bachelor's Thesis) |
|---|---|
| Supervisor name: | Bunte, K. and Dibangoye, J.S. |
| Degree programme: | Computing Science |
| Thesis type: | Bachelor's Thesis |
| Language: | English |
| Date Deposited: | 16 Jul 2024 11:06 |
| Last Modified: | 16 Jul 2024 11:06 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/33452 |
Actions (login required)
 |
View Item |