The Agent Based Intelligent Human-machine Negotiation

Negotiation is an important part of human daily social interaction,which usually involves two or more parties through dialogues.The purpose of negotiation is to eliminate disagreements between participants on the topics of negotiation so as to achieve agreements.With the continuous progress and development of computer technology,the automated negotiation technology makes it possible to create an agent as a participant to join the negotiation session against human.The research of human-agent negotiation aims at constructing sensible and skillful agents whose primary purpose is to negotiate well with humans.In order to promote the development of human-agent negotiation,the Auto-Negotiation Research Community has organized the international Automated Negotiating Agents Competition(ANAC)which always be held with the International Conference on Autonomous Agents and Multiagent Systems(AAMAS)or the International Joint Conference on Artificial Intelligence(IJCAI)annually.The problem of human-agent negotiation is still not well understood,mainly because human players are not fully rational from game theory's perspective and thus the interaction in such context is hard to model using traditional ways.To improve the performance of the agent,we propose a novel strategy for complex human-agent negotiation – that is – multiple issues,unknown opponent preferences as well as real-time constraints.This generic framework makes it possible to model opponent behaviour during negotiation session and make reasonable decisions to establish agreements with human players.The framework is able to combine with other algorithms without many changes.By combining with Gaussian Processes,Simulated Annealing and Transfer Learning techniques,we build an agent and analyze the results of extensive experiments.The results show that it is able to outperform human counterparts,in both high and low conflictive negotiation scenarios.In this thesis,we firstly introduced the background,the research status and the main idea of this thesis.Preliminaries of this thesis were introduced secondly,then we analyzed the process of human-agent negotiation and provided the generic framework for human-agent negotiation.This framework makes it possible to model opponent through the utility of agent under the counter offer sent by the human player.We built an agent with Gaussian Processes technique as model strategy and Simulated Annealing technique as offering strategy,the results of our experiments show that this agent reaches a better performance in both high and low conflictive scenarios than others.By adding the Transfer Learning technique into the framework,we eliminate the situation that the data is not enough for opponent modeling at the beginning of the negotiation.