Aristotle University Of Thessaloniki (AUTH)
Μέθοδοι ενισχυτικής μάθησης σε συστήματα πρακτόρων
Abstract
dc:descriptionThis dissertation pertains to the area of Machine Learning and especially to the subfield of Reinforcement Learning. Reinforcement learning comprises an appealing solution to problems with limited environmental feedback. The reinforcement learning framework provides the appropriate tools for solving complex problems, unlike other machine learning frameworks where correct labeled examples are necessary. For example, it is possible that the environment that an autonomous agent will act, may be unknown. Despite the research efforts and the successes in reinforcement learning, several research topics are still open. The contribution of this thesis is two-fold: a) it concerns the deployment of methods in multiagent systems and b) it uses the reinforcement learning framework to solve complex problems like focused crawling and ensemble pruning. Firstly, the thesis concentrates on the problem of coordinating a group of autonomous agents in order to achieve a common goal. For dealing with this problem, a reinforcement learning method is presented that is based on learning a set of strategies. Using a fusion procedure the individual decisions are combined, in order to follow a common strategy. Additionally, a method is presented for accelerating the learning procedure in reinforcement learning agents that can be applied in both multiagent and single agent systems. The proposed method uses multiple mapping functions, in order to transfer knowledge from a source task to a more complicated target task. Furthermore, an approach for focused crawling is proposed, which is based on the reinforcement learning framework. More specifically, the proposed approach learns a policy of selecting an appropriate classifier for link scoring during the crawling process. This is an indirect way to face the problem of focused crawling, without requiring the direct learning of selecting a link to follow or not. Finally, the problem of pruning an ensemble of classifiers is modeled as a reinforcement learning problem. The agent learns a policy of inserting or not a classifier in the ensemble. The proposed approach uses a general definition of the reward function in order to stress the capability of using different instantiations of the performance metric, the performance evaluation method depending on the requirements of the domain or the preferences of the data analyst. Additionally, a taxonomy of the existing ensemble pruning methods is proposed.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2009
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Παρτάλας, Ιωάννης
- Partalas, Ioannis
Subjects
dc:subject × 10Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/19231