Integration of Learning, Situational Power and Goal Constraints Into Time-Dependent Electronic Negotiation Agents

Abstract

In the past decade, electronic negotiation has become an important research topic in the field of information systems. A desirable goal of negotiation agents is to understand their owners' requirements, and to learn their opponents' behavior, thereby lessening the involvement of human beings. Studies on human negotiation bring out that several issues can affect a human's negotiation behavior, including learning an opponent's behavior, exerting power on an opponent, and setting an individual goal to improve the level of accomplishment. Research on incorporating these issues into negotiation agents is, however, still at an infancy state. We therefore take up this topic in this thesis. Researchers have proposed many different negotiation agents that follow a preset behavior based on human models of negotiation. In this thesis, we consider one such model, known as the time-dependent-tactical model, which is used by human negotiators and in which the values of the negotiating issues are determined based on the time elapsed in the negotiation. A learning mechanism for this model might be beneficial, because this model is frequently used in electronic negotiation. Thus, we propose heuristic algorithms that estimate the parameters of an agent's time-dependent-tactical model, and that then react to the estimated parameters for achieving higher negotiation performance. Besides learning, we incorporate two other factors that have been found to affect a human negotiation outcome. These are situational power, which represents differences in negotiators' status based on market conditions, and goal constraints, which stand for the levels of accomplishment negotiators try to strive for. To validate the impacts of learning, situational power and goal constraints in electronic negotiation, we first present how to integrate these features into negotiation agents, and then conduct simulations. With 187,500 simulation runs, we observe that our learning algorithms are effective in improving both individual and dyadic negotiation performances. For the effects of situational power and goal constraints, we obtain congruent results between human and electronic negotiations. By incorporating learning into situational power and goal constraints, we achieve significant joint effect between learning and situational power as well as that between learning and goal constraints. In summary, this thesis provides three primary contributions to the fields of information systems and electronic-commerce research. First, we have designed algorithms for learning an opponent's negotiation behavior. Second, our learning algorithms are found to be effective in improving negotiation performance. Third, we have shown how learning can be integrated with situational power and goal constraints, although this is not a major focus in this study. Finally, the agreement on the joint effects of learning, situational power and goal constraints between human and electronic negotiations suggests that our integrated design of the agent appears to be effective.