Aristotle University Of Thessaloniki (AUTH)
Μηχανισμοί ενισχυτικής μάθησης και εξελικτικής υπολογιστικής για αυτόνομους πράκτορες
Abstract
dc:descriptionThe ultimate goal of artificial intelligence is the creation of fully autonomous systems, which will be able to learn, reason, evolve and function in the real world. Such systems are usually referred to as autonomous agents. One of the most appropriate paradigms for creating autonomous agents is that of reinforcement learning. In reinforcement learning problems the goal is to find a policy, a mapping of states to actions, without examples of correct behavior, but only with positive or negative rewards based on the goal the agent is trying to achieve. The optimal policy maximizes the long-term reward of the agent. One of the main ingredients of a reinforcement learning system is the value function, a function that estimates the long-term expected reward for every state or state-action pair. For small-scale problems it can take the form of an array. For larger problems though, the function needs to be represented by a function approximator in parametric form. The reason is the generalization capabilities of the approximator, which will help the agent take correct actions for states that has not encountered before. The goal of the dissertation is the autonomous adaptation of function approximators with the use of reinforcement learning and evolutionary computing. The algorithm will adapt the parameters of the function approximator to the problem at hand with little or no human input. The name of the method developed is NEAR (NeuroEvolution of Augmented Reservoirs) and uses three basic ideas: a) Echo state networks (ESN), as function approximators, a model of reservoir computing with recursive nature and capabilities of linear learning rules and modeling non-linear environments and non-Markovian state signals, b) ΝΕΑΤ (NeuroEvolution of Augmented Topologies) methodology as a meta-search algorithm, adapted to evolve ESNs and c) evolution coupled with learning with the goal of evolving ESNs that are better able to learn. Evolution performs global-search in the space of parameters, while learning performs local-search for the final tweaking of parameters towards the global optimum. Moreover, a transfer learning procedure was tested in order to transfer knowledge in the form of reservoirs, developed in a source task to a similar target task, with the goal to improve the performance and speed of learning in the target task. The NEAR method was tested on ten (10) different reinforcement learning testbeds and four (4) time-series prediction problems in a supervised learning mode. NEAR was compared against state-of-the-art algorithms and was found superior in most of the testbeds. In addition, ESNs and NEAR were tested in three more demanding problems: in the bidding mechanisms of trading agents for supply chain management and advertisement auctions and as a mixed strategy model in the game of Poker.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2012
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Chatzidimitriou, Kyriakos
- Χατζηδημητρίου, Κυριάκος
Subjects
dc:subject × 28- Ενισχυτική μάθηση
- Εξελικτική υπολογιστική
- Νευρωνικά δίκτυα
- Δίκτυα ηχωικών καταστάσεων
- Νευροεξέλιξη
- Χρονοσειρές
- Αυτόνομοι πράκτορες
- Τεχνητή νοημοσύνη
- Reinforcement learning
- Evolutionary computing
- Neural networks
- Echo state networks
- Neuroevolution
- Time-series
- Autonomous agents
- Artificial intelligence
- Φυσικές Επιστήμες
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Κοινωνικές Επιστήμες
- Εκπαίδευση
- Natural Sciences
- Computer and Information Sciences
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
- Social Sciences
- Education
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/30602