1201 Learning Disjunctive Preferences for Negotiating Effectively * (Extended Abstract) Reyhan Aydo˘ gan Department of Computer Engineering Bo˘ gaziçi University Bebek, 34342, Istanbul,Turkey reyhan.aydogan@gmail.com Pınar Yolum Department of Computer Engineering Bo˘ gaziçi University Bebek, 34342, Istanbul,Turkey pinar.yolum@boun.edu.tr ABSTRACT Successful negotiation depends on understanding and responding to participants’ needs. Many negotiation approaches assume iden- tical needs (e.g., minimizing costs) and do not take into account other preferences of the participants. However, preferences play a crucial role in the outcome of negotiations. Accordingly, we pro- pose a negotiation framework where producer agents learn the pref- erences of consumer preferences over time and negotiates based on this new knowledge. Our proposed approach is based on inductive learning but also incorporates the idea of revision. Thus, as the ne- gotiation proceeds, a producer can revise its idea of the customer’s preferences. This enables us to learn conjunctive as well as disjunc- tive preferences. Even if the consumer’s preferences are specified in complex ways, such as conditional rules, our approach can learn and guide the producer to create well-targeted offers. Our experi- mental work shows that our proposed approach completes negoti- ation faster than similar approaches, especially if the producer will not be able to satisfy consumer’s requests properly. Categories and Subject Descriptors I.2.11 [Distributed Artificial Intelligence]: Multiagent Systems General Terms Algorithms; Experimentation Keywords Negotiation, Inductive Learning, Ontology, Semantic Similarity 1. INTRODUCTION In automated service negotiation, a consumer agent that repre- sents a real user and a provider agent that represents a business interact to reach a consensus about a service description. These agents interact by turn taking: Consumer starts the negotiation by requesting a service. If the producer cannot fulfill this need, it pro- poses a counter offer and so on. * This research is supported by Bo ˘ gaziçi University Research Fund under grant BAPA7A102 and Turkish Research and Technology Council CAREER Award under grant 105E073. Cite as: Learning Disjunctive Preferences for Negotiating Effectively (Short Paper), Reyhan Aydo˘ gan and Pınar Yolum, Proc. of 8th Int. Conf. on Autonomous Agents and Multiagent Systems (AAMAS 2009), Decker, Sichman, Sierra and Castelfranchi (eds.), May, 10–15, 2009, Bu- dapest, Hungary, pp. XXX-XXX. Copyright c  2009, International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org). All rights reserved. In order to negotiate successfully, participants need to consider each others’ service preferences and generate offers accordingly. However, preferences of participants are almost always private and hence cannot be accessed by others. The best that can happen is that participants may learn about each others’ preferences over time and through interactions. As agents learn about each others’ prefer- ences, they can provide better-targeted offers and thus enable faster negotiation. During the learning process, consumer’s requests are accepted as a positive example whereas the producer’s counter-offers that are rejected by the consumer are taken as negative examples. Ac- cording to the learning algorithm, producer’s services are filtered so that the producer selects a service as an offer from a smaller set of possible services. Learning consumer preferences during negotiation requires an incremental learning algorithm since the training data is gained dur- ing the negotiation process. For this purpose, we have previously extended the Candidate Elimination Algorithm (CEA) [2] to handle disjunctions [1]. In this work, we develop a different extension that benefits from service ontologies better and filters redundant offers earlier. 2. PROPOSED APPROACH Our proposed approach can retract its hypothesis about what it has learned as more interactions take place. Further, it uses an un- derlying ontology of service features for revising hypothesis as nec- essary. Compared to existing approaches, our proposed approach can facilitate faster negotiation of service descriptions. If no con- sensus can be found, it signals this early in the negotiation. First, according to CEA, all of the hypotheses in the most general set should cover the entire positive sample set. This rule prevents learning disjunctives since it is the union of more than one hypothe- ses and it cannot be covered by a single hypothesis with the con- dition of excluding negative samples. Therefore, in our learning algorithm we change this rule with that a positive sample should be covered by at least one of the hypotheses in the most general set. Furthermore, in some cases, a revision may be required when there is no more hypothesis in the most general set that is consistent with the incoming positive sample. In such a case, we need to add a new hypotheses covering this new positive sample while excluding all the negative samples. Hence, we require to keep the history of negative samples. Second, when a positive sample comes, generalization of the specific set is performed in a controlled way with a threshold value, Θ. As far as disjunctive concepts are concerned, there should be more than one specific hypotheses in the most specific set. Thus, Cite as: Learning Disjunctive Preferences for Negotiating Effectively, (Extended Abstract), Reyhan Aydoğan, Pınar Yolum, Proc. of 8th Int. Conf. on Autonomous Agents and Multiagent Systems (AAMAS 2009), Decker, Sichman, Sierra and Castelfranchi (eds.), May, 10–15, 2009, Budapest, Hungary, pp. 1201–1202 Copyright © 2009, International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org), All rights reserved.