Choosing Optimal Seller Based on off-line Learning Negotiation History and K-armed Bandit Problem

When a buying agent carries user’s requirements to look for a selling agent for negotiation, there may be some sellers to meet user’s requirement, but the buyer must choose a seller who can not only meet user’s requirements, but also make the buyer’s negotiation outcome gain maximal utility before negotiation. This paper solves problem of choosing seller before negotiation in order to improve accuracy of the multi-issue negotiation and buyer’s utility. In order to fully utilize negotiation history, this paper transforms the problem of choosing seller into K-armed bandit problem to solve. Several improved algorithms, which are used to learn reward distribution by off-line learning, and combine technologies for K-armed bandit problem and learning by neural network, are presented. Finally, combining the improved algorithms with trust vectors improves accuracy and practicability of choosing a seller. The experiment proves validity of the work in application.