Multi-agent Learning And Negotiation In Electronic Marketplace

More and more distributed computer systems turns to be treated as interactive agents with the development of internet. The agent oriented methodology provides us a powerful means for the analyzing, designation, realizing and evaluation of complex system in distributed environment. Multi-Agent System (MAS) is a comparatively new approach in information science and Distributed Artificial Intelligence (DAI). Since the original study in 1980s, the idea of MAS has been extensively accepted and MAS are causing more and more study interest now. The economic globalization requires every enterprise response as quickly as possible in the dynamic changing environment. It is to meet this requirement that Electronic Marketplaces (EM) is put forward. EM is a typical application of MAS in internet environment, which itself can be treated as a combination of many distributed, autonomous and interactive agents .The content of this dissertation can be concluded theoretically and practically. Theoretically, this dissertation belongs to MAS, its content covers the main problem of MAS, including multiagent learning, negotiation, inference and interaction mechanism. Practically, this dissertation discusses several key problems in EM application, including dynamic pricing, resources allocation trading negotiation. The creative work can be concluded as following.1 In the application background of dynamic pricing of sellers in electronic market, this paper put forwards an efficient online learning algorithm (IIFPQL), which integrates the observed objective actions as well as the subjective inferential intention of the opponents. Algorithm simulation of typical coordination games indicates the increase of agent's utility.2 Concerning the dynamic pricing of sellers in electronic market, this paper gives a Makov game learning model. Simulations under pricing model tested the effectiveness of the IIFPQL.3 The dissertation puts forward three multiagent cooperative learning algorithm. First, to counter the shortage of Q learning, it gives a black board model based multiagent learning algorithm (BBMML), the agents cooperatively update the Q table by a coordination function. Second, it gives E-BBMML, which explores the state-action space by evolutionary agent. Third, it gives a space-partition based multiagent cooperative learning algorithm (SSPML), which specializing each agent by a sub-space of state space.4 The dissertation puts forward a market-mechanism-based multiagent negotiation model (MMN). As a mediator. The market agent summarizing the individual information of each agent, then returns the summarized result to them. Each agent is assumed to compute its best demand by programming algorithm to maximize its utility. The paper gives the pricing algorithm and the allocation algorithm. The simulation results prove to be better than initial allocation solution. To protect the agents' privacy, the author gives a marginal utility based market mechanism...