The field of Artificial Intelligence has experienced a peak in popularity, becoming a topic of widespread discussion to the point where, even politics need a representative definition of AI. The subjectivity in defining "intelligence" has lead to different categorization of AIs, with the strongest one being between weak AI, which automates specific tasks without emulating human understanding, and strong AI, which aims to surpass human intelligence but remains theoretical. In past years, artificial intelligence primarily operated within well-defined boundaries, executing specific tasks based on programming and training, responding to the "weak AI" definition. AI's practical applications span various domains, improving safety, efficiency, and decision-making in transportation, healthcare, manufacturing, art, finance, and game playing. This thesis focuses on the domain of game playing, a longstanding area of AI research, with games offering clear rules and measurable outcomes, making them ideal test beds for algorithms and frameworks. The presence of opponents or teammates in games introduces complexity, uncertainty, and incomplete information, enabling novel and dynamic scenarios for AI research. Within the game-playing domain, the Multiplayer Online Battle Arena (MOBA) genre is of particular interest. MOBAs involve two teams of five players battling until one team's main structure is destroyed. These games emphasize complex team strategies and have gained popularity. AI's potential applications in MOBAs range from commentary generation to automatic camera control during e-sports matches. The primary objective of this thesis is to design a novel framework that leverages weak limited memory AI and deep reinforcement learning to train intelligent agents for autonomously playing MOBA games. "Derk’s Mutant Battlegrounds", through its OpenAI Gym compatible counterpart "Derk's gym", serves as the environment for this work, where each in-game character will be guided by its own intelligent agent, realized through the means of the PyTorch framework. This research aims to contribute to the advancement of AI in the context of complex gameplay scenarios, opening new possibilities for autonomous gameplay in MOBAs.
Il campo dell'Intelligenza Artificiale ha vissutoun picco di popolarità, diventando un argomento di discussione diffuso al punto che persino la politica ha bisogno di una definizione rappresentativa di IA. La soggettività nella definizione di "intelligenza" ha portato a diverse categorizzazioni delle IA, la più forte delle quali è quella tra IA debole, che automatizza compiti specifici senza emulare la comprensione umana, e IA forte, che mira a superare l'intelligenza umana ma rimane teorica. Negli anni passati, l'intelligenza artificiale operava principalmente entro confini ben definiti, eseguendo compiti specifici basati sulla programmazione e sull'addestramento, rispondendo alla definizione di "IA debole". Le applicazioni pratiche dell'IA spaziano in vari ambiti, migliorando la sicurezza, l'efficienza e il processo decisionale nei trasporti, nella sanità, nella produzione, nell'arte, nella finanza e nel game playing. Questa tesi si concentra sul dominio del game playing, un'area di ricerca sull'IA di lunga data, con giochi che offrono regole chiare e risultati misurabili, rendendoli banchi di prova ideali per algoritmi e framework. La presenza di avversari o compagni di squadra nei giochi introduce complessità, incertezza e informazioni incomplete, consentendo scenari nuovi e dinamici per la ricerca sull'IA. Nell'ambito dei giochi, il genere Multiplayer Online Battle Arena (MOBA) è di particolare interesse. I MOBA prevedono che due squadre di cinque giocatori si affrontino fino a che la struttura principale di una delle due squadre venga distrutta. Questi giochi enfatizzano complesse strategie di squadra e hanno guadagnato popolarità. Le potenziali applicazioni dell'intelligenza artificiale nei MOBA vanno dalla generazione di commenti al controllo automatico della telecamera durante le partite di e-sport. L'obiettivo primario di questa tesi è progettare un nuovo framework che sfrutti l'IA a memoria limitata debole e il deep reinforcement learning per addestrare agenti intelligenti per giocare autonomamente ai MOBA. "Derk's Mutant Battlegrounds", attraverso la sua controparte compatibile con OpenAI Gym "Derk's gym", funge da ambiente per questo lavoro, dove ogni personaggio in-game sarà guidato da un proprio agente intelligente, realizzato attraverso il framework PyTorch. Questa ricerca vuole contribuire al progresso dell'IA nel contesto di scenari di gioco complessi, aprendo nuove possibilità di gioco autonomo nei MOBA.
DEEP DETERMINISTIC POLICY GRADIENT PER L'ADDESTRAMENTO DI AGENTI IN GIOCHI MOBA
SERVEDIO, GIOVANNI
2022/2023
Abstract
The field of Artificial Intelligence has experienced a peak in popularity, becoming a topic of widespread discussion to the point where, even politics need a representative definition of AI. The subjectivity in defining "intelligence" has lead to different categorization of AIs, with the strongest one being between weak AI, which automates specific tasks without emulating human understanding, and strong AI, which aims to surpass human intelligence but remains theoretical. In past years, artificial intelligence primarily operated within well-defined boundaries, executing specific tasks based on programming and training, responding to the "weak AI" definition. AI's practical applications span various domains, improving safety, efficiency, and decision-making in transportation, healthcare, manufacturing, art, finance, and game playing. This thesis focuses on the domain of game playing, a longstanding area of AI research, with games offering clear rules and measurable outcomes, making them ideal test beds for algorithms and frameworks. The presence of opponents or teammates in games introduces complexity, uncertainty, and incomplete information, enabling novel and dynamic scenarios for AI research. Within the game-playing domain, the Multiplayer Online Battle Arena (MOBA) genre is of particular interest. MOBAs involve two teams of five players battling until one team's main structure is destroyed. These games emphasize complex team strategies and have gained popularity. AI's potential applications in MOBAs range from commentary generation to automatic camera control during e-sports matches. The primary objective of this thesis is to design a novel framework that leverages weak limited memory AI and deep reinforcement learning to train intelligent agents for autonomously playing MOBA games. "Derk’s Mutant Battlegrounds", through its OpenAI Gym compatible counterpart "Derk's gym", serves as the environment for this work, where each in-game character will be guided by its own intelligent agent, realized through the means of the PyTorch framework. This research aims to contribute to the advancement of AI in the context of complex gameplay scenarios, opening new possibilities for autonomous gameplay in MOBAs.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_Servedio-Esse3.pdf
File consultabile da postazioni biblioteca
Descrizione: Tesi di Laurea Magistrale
Dimensione
4.31 MB
Formato
Adobe PDF
|
4.31 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14253/80985