Type:	Monografia
Title:	Persistent Monte Carlo Tree Search for agents operating in complex environments: a case study on Citadels
Author:	Camilotto, Andrei Carlesso Bonelli, Djonatan Riquelme Clein Fiorentin, Eduardo Vinicius Perissinotto Santos, João Luís Almeida
First advisor:	Grando, Felipe
Resume:	Decision-making in multiplayer, partially observable games poses significant challenges due to hidden information and large branching factors. This study introduces a persistent Monte Carlo Tree Search (MCTS) framework designed to operate efficiently in such environments, using the strategic board game Citadels as a case study. The key contribution is the decoupling of the tree construction phase from the application phase: decision trees are generated offline through environment simulations, and relevant statistics are stored in tabular form, enabling rapid querying during execution without the need for additional simulations. Experimental evaluation demonstrates that exposure to diverse opponents during training produces models with robust and generalizable strategies, capable of strong performance across a wide range of game scenarios. In particular, models trained in an environment where opponents were stochastically sampled consistently outperformed those trained against single opponent types and exhibited stable performance across training runs. These findings underscore the effectiveness of combining MCTS with persistent knowledge storage to produce fast, reli- able, and adaptable agents in complex domains.
Keywords:	Inteligência artificial Algoritmos Método de Monte Carlo Jogos de tabuleiro
Language:	por
Country:	Brasil
Publisher:	Universidade Federal da Fronteira Sul
Acronym of the institution:	UFFS
College, Institute or Department:	Campus Chapecó
Type of Access:	Acesso Embargado
URI:	https://rd.uffs.edu.br/handle/prefix/9196
Issue Date:	2025
Appears in Collections:	Ciência da Computação

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.