Unit:
Κατεύθυνση Φιλοσοφία των Επιστημών και της ΤεχνολογίαςLibrary of the School of Science
Author:
Thomas Konstantinos
Supervisors info:
Αριστείδης Χατζής, Αναπληρωτής Καθηγητής, Τμήμα Ιστορίας και Φιλοσοφίας της Επιστήμης, Εθνικό και Καποδιστριακό Πανεπιστήμιο Αθηνών
Γιούλη Φωκά-Καβαλιεράκη, Διδάκτωρ, Τμήμα Ιστορίας και Φιλοσοφίας της Επιστήμης, Εθνικό και Καποδιστριακό Πανεπιστήμιο Αθηνών
Χρυσόστομος Μαντζαβίνος, Καθηγητής, Τμήμα Ιστορίας και Φιλοσοφίας της Επιστήμης, Εθνικό και Καποδιστριακό Πανεπιστήμιο Αθηνών
Original Title:
Building a Reinforcement Learning A.I. for the Iterated Prisoner's Dilemma using Soar cognitive architecture
Translated title:
Building a Reinforcement Learning A.I. for the Iterated Prisoner's Dilemma using Soar cognitive architecture
Summary:
This thesis involves the creation of an Artificial Intelligence agent that uses Reinforcement Learning (Q-Learning) in order to discern an optimal strategy to the game-theoretic Iterated Prisoner's Dilemma game. The agent is built using the Soar cognitive architecture and starts with sole knowledge of its two possible moves - Cooperation and Defection. Exploring the problem space, blindly at first by executing random moves, the A.I. agent quickly develops an intuition as of how the game is played through the Rewards it receives after each move. As more rounds go by, the agent starts realizing ways to handle the different opponents in order to maximize its payoff. Once its training is complete we examine how it fares compared to the top deterministic IPD strategy, Tit for Tat, which was the winner of both Axelrod’s tournaments. Eventually, we examine the aggregate preference that the agent heuristically developed, for each individual move and note how they contrast with the rules of known deterministic strategies.
Main subject category:
Technology - Computer science
Keywords:
Soar, Cognitive Architecture, Iterated Prisoner's Dilemma, Artificial Intelligence, Machine learning, Reinforcement Learning
