Machine Learning Approach for Optimizing Negotiation Agents

• Main Author: Ng, Sok Choo
• Format: Thesis
• Language: English
• Published: 2007
• Online Access: http://psasir.upm.edu.my/id/eprint/5209/1/FSKTM_2007_12.pdf; http://psasir.upm.edu.my/id/eprint/5209/
• Institution: Universiti Putra Malaysia

The increasing popularity of Internet and World Wide Web (WWW) fuels the rise of electronic commerce (E-Commerce). Negotiation plays an important role in ecommerce as business deals are often made through some kind of negotiations. Negotiation is the process of resolving conflicts among parties having different criteria so that they can reach an agreement in which all their constraints are satisfied. Automating negotiation can save human’s time and effort to solve these combinatorial problems. Intelligent Trading Agency (ITA) is an automated agentbased one-to-many negotiation framework which is incorporated by several one-toone negotiations. ITA uses constraint satisfaction approach to evaluate and generate offers during the negotiation. This one-to-many negotiation model in e-commerce retail has advantages in terms of customizability, scalability, reusability and robustness. Since negotiation agents practice predefined negotiation strategies, decisions of the agents to select the best course of action do not take the dynamics of negotiation into consideration. The lack of knowledge capturing between agents during the negotiation causes the inefficiency of negotiation while the final outcomes obtained are probably sub-optimal. The objective of this research is to implement machine learning approach that allows agents to reuse their negotiation experience to improve the final outcomes of one-to-many negotiation. The preliminary research on automated negotiation agents utilizes case-based reasoning, Bayesian learning and evolutionary approach to learn the negotiation. The geneticbased and Bayesian learning model of multi-attribute one-to-many negotiation, namely GA Improved-ITA and Bayes Improved-ITA are proposed. In these models, agents learn the negotiation by capturing their opponent’s preferences and constraints. The two models are tested in randomly generated negotiation problems to observe their performance in negotiation learning. The learnability of GA Improved-ITA enables the agents to identify their opponent’s preferable negotiation issues. Bayes Improved-ITA agents model their opponent’s utility structure by employing Bayesian belief updating process. Results from the experimental work indicate that it is promising to employ machine learning approach in negotiation problems. GA Improved-ITA and Bayes Improved-ITA have achieved better performance in terms of negotiation payoff, negotiation cost and justification of negotiation decision in comparison with ITA. The joint utility of GA Improved-ITA and Bayes Improved-ITA is 137.5% and 125% higher than the joint utility of ITA while the negotiation cost of GA Improved-ITA is 28.6% lower than ITA. The negotiation successful rate of GA Improved-ITA and Bayes Improved-ITA is 10.2% and 37.12% higher than ITA. By having knowledge of opponent’s preferences and constraints, negotiation agents can obtain more optimal outcomes. As a conclusion, the adaptive nature of agents will increase the fitness of autonomous agents in the dynamic electronic market rather than practicing the sophisticated negotiation strategies. As future work, the GA and Bayes Improved-ITA can be integrated with grid concept to allocate and acquire resource among cross-platform agents during negotiation.