Research On Bilateral Multi-issue Automated Negotiation

In a competitive, open e-commerce environment, negotiation has become the inevitable choice to resolve conflicts of interest. With Agent technology and multi-Agent system development and gradual maturation, automated negotiation will further promote the development of electronic commerce. Automated negotiation research includes three aspects:protocol, issue and negotiation strategy. Negotiation strategy research has made a lot of valuable research in recent years, the machine learning method is introduced into the decision-making model in this area become a new hotspot.Today, automated negotiation has been a transition period from theory to application, and introduction of machine learning algorithms in negotiation decision model need to be further improved to meet the needs of practical problems. Starting with an introduction to the foundational negotiation theory and an analysis of the automated negotiation, based on the bilateral multi-issue automated negotiation framework and negotiation decision model research, and aiming at the weak points of machine learning algorithms based negotiation decision model, the thesis proposes new negotiation decision models to improve the negotiation performance. The main research in this thesis can be classed as follows:1. A bilateral multi-issue automated negotiation framework is presented. The paper gives a framework definition for bilateral multi-issue negotiation, set the detailed negotiation participant,issue,protocol,decision model. In the negotiation decision model, we took into account the computational complexity of generating counter-offer, the number of negotiation rounds, one's own utility and overall effectiveness of the two sides negotiate to measure the perfonnance of decision model, and we design four kinds of negotiation decision model to deal with different negotiation environment and application requirements.2. A learning opponent's utility function negotiation decision model is presented. By the use of transitive support vector machine, the negotiation model learns opponent's approximate utility function from the history data's proposal feature, and then considering the effectiveness of the trade-off between opponent and one's own utility to get the optimal solution. The model overcomes the lack of marked training samples (more unmarked data and less marked data). Experimental results show that the model is effective and efficient in the environment where lack of marked training samples.3. A negotiation decision model based on learning opponent's preference is presented. In many cases the multiple issues negotiation have more chance to achieve win-win contracts. However, effective and efficient multi-issue negotiation requires an agent to have some indication of its opponent's preferences over these issues, and in competitive domains, such as e-commerce, an agent will not reveal this information. This model learns opponent's preferences from negotiation history data, and then weighs the preferences of both issues to calculate the specific value of the counter-offer. This model avoids the classical model's assumption that available of some opponent's private information, experiments show that this model can effectively reduce negotiation time and increase the joint utility.4. A negotiation decision model based on learning bilateral negotiation track is presented. Some of the current research on decision model only to learn opponent's negotiation track. When establish the relationship between opponent's preference and negotiation track, these researches not taking into account the one's own negotiation behavior can have some impact of the rival's negotiation behavior. But negotiation process is a bilateral interaction; one's own values and the concessions on a certain issue could have a direct impact on the opponent's offer, and then appears in the opponent's negotiation track. In view of this situation, we proposed a negotiation decision model based on learning bilateral negotiation track. Experiments showed that: comparing the current studies, learning bilateral negotiation track can have greater flexibility and adaptability, the establishment of relation between the bilateral negotiation track and opponent's preference is more in line with the actual negotiation situation. 5. A negotiation decision model based on ensemble learning is presented. At present studies, automated negotiation is the introduction of a single machine learning method. But a single machine learning method's classification is weak, and unable to get a better negotiation performance, when lack of training samples (only a small amount of negotiation history data). So we proposed a negotiation decision model based on ensemble learning. First, this model calculates the Mahalanobis distance in feature space of training samples. With the use of KNN algorithm, the model learns opponent's approximate preference. And then gradually narrow the prediction values and actual values of opponent's preference through ensemble learning. Experimental results show that the model is effective and efficient in the environment where lack of training samples.