Multi-Issue negotiation between Agents is a complicated course in which negotiating Agents mutually exchange offers. Solving the problem of choosing seller before negotiation has important practical value in e-commerce. The problem is solved in this paper to improve accuracy of the multi-issue negotiation and buying Agent’s utility. In order to fully utilize negotiation history, tradeoff exploration and exploitation, the problem of choosing seller is transformed into a K-armed bandit problem. A model for measuring trust and reputation is presented, several improved algorithms, which are used to learn reward distribution and combine learning with technologies for K-armed bandit problem, are presented. Finally, the combination of the improved algorithms, the trust and reputation improves the accuracy and practicability of choosing a selling Agent. Several experiments prove validity of the work in application.