Beliefs Learning in Fuzzy Constraint-directed Agent Negotiation
| Autor: | Ting-Jung Yu, 余丁榮 |
|---|---|
| Rok vydání: | 2009 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 97 This dissertation presents a framework for learning opponent’s beliefs to improve the quality of negotiation and the overall competence of each agent in a multi-agent problem-solving environment. We first present a fuzzy constraint-directed approach to agent negotiation, including basic definitions, behavior model, communication messages, and computational model of negotiation. Then, this negotiation framework is extended to incorporate a learning element for capturing opponent’s beliefs. Through exchange messages, agent can deduce the features of opponent’s (1) preference functions, (2) the importance of issues, and (3) negotiation strategy and to predict opponent’s behavior state. Additionally, we also construct an instance repository for matching the proximate instances to predict opponent’s next feasible proposals and to speed up the convergence of the expected negotiation goal. For beliefs learning, two learning models are proposed, Bayesian learning and fuzzy probability learning, for different negotiation environments, The results of numerical experiments including the speed of convergence and the quality of negotiation reveal that the Bayesian learning method fairs better at regularizing the opponent’s behavior models in a long-term stable environment, while the fuzzy probability learning approach is more adept for approximating the opponent’s behavior patterns in short-term volatile situation. Finally, to demonstrate the practicality of the proposed approach, two real-world example applications, a bi-lateral multi-issue negotiation for insurance policy and a multi-lateral negotiation for supply chain, are successfully implemented and also exhibit the benefits of learning-enable agent negotiation. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|