This paper explores the design of artificial intelligence for collectible card games, using *Dataclysm* — a digital title in development by the Pet The Dog team — as a case study. After a review of the state of the art in perfect and imperfect information games, the main AI methods applicable to CCGs are analyzed — from Monte Carlo Tree Search to Deep Reinforcement Learning, and hybrid approaches involving neural networks and self-play — highlighting their advantages, limitations, and adaptability to the specific mechanics of *Dataclysm*. The game is described in its key elements (game state, card types, and turn phases), showing how, thanks to a limited number of cards and mechanics as well as a constrained branching factor, it is possible to perform fast simulations and implement a computationally intensive MCTS in real time. The main contribution consists of a comparative evaluation of the three AI approaches, followed by a set of concrete implementation proposals: pure MCTS for lightweight prototypes, Deep RL for optimal policies at the cost of high computational resources, and hybrid architectures combining efficient exploration with adaptive heuristic evaluation. Finally, the need for pruning techniques, rapid retraining after each card update, and reward shaping is discussed to ensure robustness and generalization of the AI in a dynamic context. Part of the writing (specifically Chapters II and IV) was carried out jointly with Filippo Barboni, who provided a fundamental contribution.
Questo elaborato affronta la progettazione di un’intelligenza artificiale per giochi di carte collezionabili, prendendo come caso di studio Dataclysm, un titolo digitale in sviluppo dal team Pet The Dog. Dopo una ricostruzione dello stato dell’arte nei giochi a informazione perfetta e imperfetta, vengono analizzati i principali metodi di IA applicabili ai CCG — dalla ricerca Monte Carlo Tree Search al Deep Reinforcement Learning, fino ad approcci ibridi con reti neurali e self-play — mettendone in luce vantaggi, limiti e adattabilità alle meccaniche specifiche di Dataclysm . Il gioco è descritto nei suoi elementi chiave (stato di gioco, tipologia di carte e fasi turnali) e si mostra come, grazie a un numero contenuto di carte e meccaniche nonché a un branching factor limitato, sia possibile realizzare simulazioni rapide e un MCTS “pesante” in tempo reale. Il contributo principale consiste in una valutazione comparativa dei tre approcci di IA, seguita da una proposta di implementazioni concrete: MCTS puro per prototipi leggeri, Deep RL per policy ottimali a costo di risorse computazionali elevate, e architetture ibride che combinano esplorazione efficiente e valutazione euristica adattiva. Infine, si discute la necessità di tecniche di pruning, riaddestramento rapido dopo ogni aggiornamento di carte e reward shaping per garantire robustezza e generalizzazione dell’IA in un contesto dinamico. Parte del lavoro di stesura (in particolare i Capitoli II e IV) è stato svolto congiuntamente a Filippo Barboni, a cui si deve un fondamentale apporto.
Progettazione Di Un’IA Per L’ottimizzazione Delle Strategie Nei Ccg: Caso Studio Dataclysm
RUSSO, GIOVANNI
2024/2025
Abstract
This paper explores the design of artificial intelligence for collectible card games, using *Dataclysm* — a digital title in development by the Pet The Dog team — as a case study. After a review of the state of the art in perfect and imperfect information games, the main AI methods applicable to CCGs are analyzed — from Monte Carlo Tree Search to Deep Reinforcement Learning, and hybrid approaches involving neural networks and self-play — highlighting their advantages, limitations, and adaptability to the specific mechanics of *Dataclysm*. The game is described in its key elements (game state, card types, and turn phases), showing how, thanks to a limited number of cards and mechanics as well as a constrained branching factor, it is possible to perform fast simulations and implement a computationally intensive MCTS in real time. The main contribution consists of a comparative evaluation of the three AI approaches, followed by a set of concrete implementation proposals: pure MCTS for lightweight prototypes, Deep RL for optimal policies at the cost of high computational resources, and hybrid architectures combining efficient exploration with adaptive heuristic evaluation. Finally, the need for pruning techniques, rapid retraining after each card update, and reward shaping is discussed to ensure robustness and generalization of the AI in a dynamic context. Part of the writing (specifically Chapters II and IV) was carried out jointly with Filippo Barboni, who provided a fundamental contribution.File | Dimensione | Formato | |
---|---|---|---|
TESI DI LAUREA.pdf
accesso aperto
Descrizione: File della tesi di laurea
Dimensione
1.77 MB
Formato
Adobe PDF
|
1.77 MB | 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/20.500.12075/22156