This thesis addresses the design of an artificial intelligence system for collectible card games (CCG), with particular reference to the game Dataclysm, currently under development. The work begins with an analysis of the state of the art, examining the main techniques developed for games with perfect and imperfect information, highlighting the specific challenges that arise in the context of CCGs, which are characterized by randomness, hidden information, and high decision-making complexity. To build a solid foundation for analysis and design, considerable space is dedicated to the discussion of fundamental theoretical concepts. In particular, the thesis explores topics related to intelligent agent theory, decision theory, decision-making processes under uncertainty and sequentiality, and game theory. This theoretical framework proves essential for understanding the difficulties in applying classical search and planning methods to imperfect information games like CCGs. In the applied section, a first AI prototype is described, developed through a rule-based and greedy approach, implemented in Prolog and integrated with Unity. While functional, this model shows significant limitations in terms of strategic capabilities, lack of long-term planning, and inability to dynamically adapt to the game context. To overcome these limitations, the thesis analyzes and compares three main approaches: Monte Carlo Tree Search (MCTS), deep reinforcement learning (Deep RL), and hybrid approaches based on self-play combined with neural networks. The theoretical results show that combining MCTS with machine learning models—an approach inspired by systems like AlphaZero—offers a particularly effective solution to balance exploration, heuristic evaluation, and adaptability. The work concludes with a proposal for possible future developments, including the implementation of a self-play environment, the adoption of techniques for dynamic data generation and state space abstraction, as well as experimental validation through automated matches and games against human players. This thesis represents a methodological contribution aimed at providing a useful roadmap for the evolution of intelligent agents in the field of complex digital games. Part of the writing (specifically Chapters II and V) was carried out jointly with Giovanni Russo, whose contribution was fundamental.
Questa tesi affronta il tema della progettazione di un’intelligenza artificiale per giochi di carte collezionabili (CCG), con particolare riferimento al gioco Dataclysm, attualmente in fase di sviluppo. L’elaborato parte dall’analisi dello stato dell’arte, esaminando le principali tecniche sviluppate per giochi a informazione perfetta e imperfetta, evidenziando le sfide specifiche che emergono nel contesto dei CCG, caratterizzati da casualità, informazione nascosta e una elevata complessità decisionale. Per costruire una base solida per l’analisi e la progettazione, è stato dedicato ampio spazio alla trattazione delle nozioni teoriche fondamentali. In particolare, vengono esplorati i concetti legati alla teoria degli agenti intelligenti, alla teoria delle decisioni, ai processi decisionali in condizioni di incertezza e sequenzialità, fino alla teoria dei giochi. Questo quadro teorico si rivela essenziale per comprendere le difficoltà nell’applicazione dei metodi classici di ricerca e pianificazione ai giochi a informazione imperfetta come i CCG. Nella parte applicativa, viene descritto un primo prototipo di IA sviluppato tramite un approccio rule-based e greedy, basato su Prolog e integrato con Unity. Questo modello, pur funzionante, presenta limiti significativi in termini di capacità strategica, assenza di pianificazione a lungo termine e incapacità di adattarsi dinamicamente al contesto di gioco. Per superare tali criticità, la tesi analizza e confronta tre approcci principali: il Monte Carlo Tree Search (MCTS), l’apprendimento per rinforzo profondo (Deep Reinforcement Learning) e gli approcci ibridi basati su self-play e reti neurali. I risultati teorici mostrano come la combinazione tra MCTS e modelli di apprendimento automatico, ispirata a sistemi come AlphaZero, rappresenti una soluzione particolarmente efficace per bilanciare esplorazione, valutazione euristica e adattabilità. Il lavoro si conclude con la proposta di possibili sviluppi futuri, tra cui la realizzazione di un ambiente di self-play, l’adozione di tecniche di generazione dinamica dei dati e di astrazione dello spazio degli stati, oltre alla validazione sperimentale tramite match automatici e partite contro giocatori umani. Questo elaborato si configura come un contributo metodologico volto a fornire una roadmap utile per l’evoluzione degli agenti intelligenti nell’ambito dei giochi digitali complessi. Parte del lavoro di stesura (in particolare i Capitoli II e V) è stato svolto congiuntamente a Giovanni Russo, a cui si deve un fondamentale apporto.
Progettazione Di Un’IA Per L’ottimizzazione Delle Strategie Nei CCG: Studi teorici
BARBONI, FILIPPO
2024/2025
Abstract
This thesis addresses the design of an artificial intelligence system for collectible card games (CCG), with particular reference to the game Dataclysm, currently under development. The work begins with an analysis of the state of the art, examining the main techniques developed for games with perfect and imperfect information, highlighting the specific challenges that arise in the context of CCGs, which are characterized by randomness, hidden information, and high decision-making complexity. To build a solid foundation for analysis and design, considerable space is dedicated to the discussion of fundamental theoretical concepts. In particular, the thesis explores topics related to intelligent agent theory, decision theory, decision-making processes under uncertainty and sequentiality, and game theory. This theoretical framework proves essential for understanding the difficulties in applying classical search and planning methods to imperfect information games like CCGs. In the applied section, a first AI prototype is described, developed through a rule-based and greedy approach, implemented in Prolog and integrated with Unity. While functional, this model shows significant limitations in terms of strategic capabilities, lack of long-term planning, and inability to dynamically adapt to the game context. To overcome these limitations, the thesis analyzes and compares three main approaches: Monte Carlo Tree Search (MCTS), deep reinforcement learning (Deep RL), and hybrid approaches based on self-play combined with neural networks. The theoretical results show that combining MCTS with machine learning models—an approach inspired by systems like AlphaZero—offers a particularly effective solution to balance exploration, heuristic evaluation, and adaptability. The work concludes with a proposal for possible future developments, including the implementation of a self-play environment, the adoption of techniques for dynamic data generation and state space abstraction, as well as experimental validation through automated matches and games against human players. This thesis represents a methodological contribution aimed at providing a useful roadmap for the evolution of intelligent agents in the field of complex digital games. Part of the writing (specifically Chapters II and V) was carried out jointly with Giovanni Russo, whose contribution was fundamental.File | Dimensione | Formato | |
---|---|---|---|
TESI DI LAUREA - Barboni.pdf
accesso aperto
Descrizione: Tesi di laurea
Dimensione
905.33 kB
Formato
Adobe PDF
|
905.33 kB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.12075/22153