Incorporating Feature Penalty in Reinforcement Learning for Ludo Game

Author First name, Last name, Institution

A. Tubaishat, Zayed University
S. Anwar, Institute of Management Sciences Peshawar
F. Al-Obeidat, Zayed University
B. Shah, Zayed UniversityFollow
S. Razzaq, University of Sargodha

Document Type

Conference Proceeding

Source of Publication

2023 7th IEEE Congress on Information Science and Technology (CiSt)

Publication Date

12-22-2023

Abstract

Modeling cognitive behavior in AI games has got much attention in recent years. One popular and most commonly used cognitive learning technique is Reinforcement Learning (RL). It is a self-learning ability that empowers game agents to flatter as an autodidact. RL has also got much attention in learning board games. Ludo is one of the popular board games of the Asian subcontinent. Researchers primarily focused on the reward feature of RL. This varies between 0 and 1 (i.e. No reward to maximum reward). However, there is no direct concept of penalty in RL. Focus on penalties leads to efficient learning. Such that, considering the penalty before action selection may decrease the worth of actions leading to the penalty. Keeping this in view, we have proposed a Q-learning-based Intelligent Ludo Agent (ILA) that incorporates both rewards and penalties. Q-Learning is an off-policy RL technique used in Machine Learning (ML). The information employed for learning is gained from Q-learning tables maintaining the history of rewards and penalties gained over the prior iterations. ILA has been trained using two different Q-learning strategies. The first one uses board indices with penalty information for moving a piece. Other uses situation on the board to select the appropriate strategy for moving a piece. Also for evaluating performance various matches of ILA were conducted against humans and random machine players. Human players facilitated in flourishing learning abilities of ILA. Whereas, experimentation revealed comparable learning abilities of autodidact agents. Thereby, moving a step ahead from formerly adopted strategies.

DOI Link

10.1109/cist56084.2023.10409945

ISBN

978-1-6654-6133-7

Publisher

IEEE

Volume

00

First Page

524

Last Page

530

Disciplines

Computer Sciences

Keywords

Reinforcement Learning, Ludo Game, Q-learning, Intelligent Ludo Agent, Machine Learning

Recommended Citation

Tubaishat, A.; Anwar, S.; Al-Obeidat, F.; Shah, B.; and Razzaq, S., "Incorporating Feature Penalty in Reinforcement Learning for Ludo Game" (2023). All Works. 6379.
https://zuscholars.zu.ac.ae/works/6379

Indexed in Scopus

no

Open Access

no