E-Commerce Decision Model Based on Auto-Learning

Introduction

With the rapid development of multi-agent technology, the application of agent technology in the field of e-commerce has become more and more extensive. Agent can help human users deal with complicated business activities more efficiently. In the e-commerce activities, auto-learning is the most core part, therefore, it has become the key content of e-commerce based on Agent. The research of auto-learning mainly includes three aspects (Hsu et al., 2015): 1) Agreement of Negotiation (Du et al., 2007): The rule set that the participants of the negotiation must comply with in the interaction. 2) Content of Negotiation (Tremel et al., 2016): The composition of the issues required for negotiation and the value region of all the issues under negotiation, etc. 3) Negotiation strategy (Fatima et al., 2007, Mok et al., 2005, Rahwan et al., 2007): Under the specific agreement of negotiation, how the negotiation agent selects the action of negotiation to be executed according to the current status information. This paper mainly focuses on the research of the strategy of negotiation under the bilateral multi-issue agreement (Águila et al., 2015, Gerding et al., 2006).

The study of negotiation strategy has made a lot of valuable research results. Especially in recent years, the auto-learning method has been introduced into the decision model, which has become a new research hotspot in this field. Bayesian learning (Morente-Molinera et al., 2015) may be the most widely applied learning method in the negotiation decision models. In this method, the Agent first needs to establish an opponent's preference or set of hypotheses, and set the initial probability of these assumptions; in the process of the negotiation, according to the opponent's negotiation, using Bayesian rules Update the probability of these assumptions; and finally make use of the probability distribution of these assumptions to make the negotiation decision. The evolutionary computing method encodes the strategy into genes (Gerding et al., 2006), which are constantly evolved by mutation, crossover and selection, etc., in order to arrive at an optimal strategy. The Kernel Density Estimation (referred to as KDE for short) method (Coehoorn et al., 2004) first estimates the weight of each item in the opponent's negotiation, and then calculates the counter negotiation which is most similar to the opponent's suggestion according to the similarity function and the weight of the opponent's item. The domestic research on the negotiation strategy also made a lot of achievements. For example, in literature (Stern et al., 2015), the problem of selection of sales agent is transformed into K-armed bandit problem, and the measurement model of trust and reputation is put forward. Combined with the K-arm gambling machine problem solving technique, with the application of the learning mechanism, several improved algorithms for determining the distribution of reward are proposed; based on the simulation process, the improved algorithm, trust and reputation are organically combined to select the sales agent. In literature (Gao et al., 2006), a multi-issue negotiation model (MN) is proposed, and on this basis an accelerating chaotic evolutionary algorithm (ACEA) is put forward. ACEA algorithm introduces the chaos mechanism into evolutionary computing, and adopts the compression technique to accelerate the algorithm.