Institutional Repository of Vilnius University
Neural network guided monte carlo tree search for liar's dice /
Abstract
dc:descriptionThis thesis investigates neural network guided Monte Carlo Tree Search (MCTS) for Liar's Dice. The game is a four-player imperfect-information game where players make bids while observing only their own dice and reasoning about their opponents' hidden dice. The aim of the thesis was to implement and evaluate a neural-guided Information Set Monte Carlo Tree Search (ISMCTS) system and determine whether learned action priors improve playing strength. The work implements a Liar's Dice environment, a heuristic baseline, several non-neural ISMCTS agents, and two policy networks. The policy networks were first trained using supervised learning using generated search targets and then refined through self-play learning. The experiments show that plain ISMCTS without search guidance is weak, while probabilistic action priors substantially improve performance. Learned action priors improve performance further, and self-play learning improves both policy networks. Although the network with the more complex architecture lead to the highest final win rate, its advantage over the simple network was small compared with its much higher computational cost.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Jonkus, Martynas,
Subjects
dc:subject × 1Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- eng
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308337175&prefLang=en_US