Elsevier

Computer Networks

Volume 37, Issue 2, October 2001, Pages 195-204
Computer Networks

A genetic agent-based negotiation system

https://doi.org/10.1016/S1389-1286(01)00215-8Get rights and content

Abstract

Automated negotiation has become increasingly important since the advent of electronic commerce. Nowadays, goods are no longer necessarily traded at a fixed price, and instead buyers and sellers negotiate among themselves to reach a deal that maximizes the payoffs of both parties. In this paper, a genetic agent-based model for bilateral, multi-issue negotiation is studied. The negotiation agent employs genetic algorithms and attempts to learn its opponent's preferences according to the history of the counter-offers based upon stochastic approximation. We also consider two types of agents: level-0 agents are only concerned with their own interest while level-1 agents consider also their opponents' utility. Our goal is to develop an automated negotiator that guides the negotiation process so as to maximize both parties' payoff.

Introduction

The advent of electronic commerce has revolutionized the way that contemporary business operates. Nowadays, online purchasing in becoming widely adopted and its selling volume is continuously increasing.1 Not only does electronic commerce reduces the operation cost of a business, but also makes a company's service available to customers virtually 24 h a day. In addition, goods are no longer necessarily traded at a fixed price but based on the market demand (e.g., ebay2 and priceline3). In the near future, buyers and sellers can negotiate among themselves so as to reach a deal that maximizes the payoffs to both parties. It is even more interesting to develop autonomous agents that can strive for the best deal on behalf of the traders. In this paper, we study how such an autonomous negotiation system can be built based upon agent technologies and genetic algorithms.

Negotiation [11], [13] is a process of reaching an agreement on the terms (such as price and quantity) of a transaction for two or more parties. The negotiation process typically goes through a number of iterations, and in each of which one of the parties proposes an offer and sees whether the others accept. If not, other parties can propose their counter-offers and the process repeats until a consensus is reached. For an effective negotiation, it is important that all parties are willing to concede such that the differences among themselves reduce at each round and eventually converge to an agreement. On the other hand, negotiation can also be viewed as a search process in which conflicts among different parties are resolved by finding a feasible alternative. However, negotiation is not just a matter of finding an acceptable deal, but an attempt to maximize all parties' payoffs. In order to achieve this goal, one needs to know the others' utility function. However, the utility functions are usually private and sensitive information. In this paper, we consider how this utility function can be estimated from the history of the opponent's offers.

Negotiation is common in conventional commerce, especially when large and complex business transactions are involved. Negotiation on various terms of a deal offers the potential to yield the involved parties the best payoffs. It also allows the terms of the deal (e.g., price) to be set according to the market demand and supply. However, human-based negotiation could be costly and non-optimal. Automated negotiation is therefore particularly useful due to its relatively low cost. Nevertheless, most existing e-commerce sites still employ fixed-pricing models, or allow only one-side negotiation (e.g., auction sites).

While negotiation is often beneficial to the participants, there are impediments to apply it in conventional business. The first concern is the time involved. Negotiation is often a time-consuming process because all parties desire to maximize their own payoff while they may have opposing goals. If some of the parties do not concede, it could take forever to reach an agreement. The second is that negotiation requires skillful tactics, and could be difficult for average dealers to bargain effectively on their own. The third difficulty is that all parties must first get together so as to negotiate. This imposes some restrictions on the customers since e-commerce can be worldwide and often involves people from various time zones.

There exist several types of negotiation models [8]. In this paper, a bilateral, multi-issue negotiation model is studied. Multi-issue negotiation is concerned with reaching an agreement on a deal with multiple terms. The involved parties typically do not want to reveal their underlying utility function and attempt to strive for the best deal through the negotiation process. An important topic for multi-issue negotiation is the question on how to avoid a sub-optimal deal; namely, how to avoid an agreement in which one party can modify to obtain better payoff without sacrificing the others. This is also known as pareto-inferior agreement or the problem of “leaving money on the table” [12].

There have been several approaches to automated negotiation [19]. For multi-issue negotiation, the search space is typically complex and large, and with little information on its structure apriori. One recent research direction is to address such difficult problems with genetic algorithms (GAs) (e.g., [10], [15]). GA [4] is particularly suitable for such tasks due to its efficient searching in a complex and large search space. In Oliver's approach, negotiation strategies are based on simple sequential rules with utility threshold for delimiters. One shortcoming for such representation is the lack of expressiveness. In order to enhance the strategy models, Tu et al. [15] deploy finite state machines for representing simple decision rules. In this paper, we are particularly concerned with learning issues; namely, how to speed up the negotiation by considering the offer history and adaptive mutation rate.

Personal software agents [6], [18] are continuously running programs that intimately work with their owners. Ideally, software agents should be proactive, intelligent, capable of understanding their owners' requirements, and therefore can perform tasks on behalf of their owners. While the current status is still far from its ideal goal, agent technologies have been successfully applied in various domains such as information filtering and job matching [2]. It is natural to extend the agent technology to automated negotiation tasks. Imagine a picture of our daily life in the near future. Everyone owns at least one agent as a personal assistant. These agents are responsible for various types of personal tasks, ranging from reminding appointments, to handling payments, to scheduling meetings. One task particularly about interesting to electronic business is how agents can be used to deal with online purchasing. Unlike the way we purchase today, agents may negotiate with the sellers on the price among other terms, and even aligns with other buyer agents to bargain for a better deal. Agent negotiation involves three major steps: initiating a negotiation task by proposing an offer, evaluating the opponent's proposal if the offer is not accepted, and suggesting a new counter-proposal. This process is repeated several rounds until an agreement is reached.

An ideal negotiation system should minimize the involvement of human beings. As noted earlier, software agents are capable of undertaking the task. One reason is that personal agents have the access to their owners' schedule, and thus can infer a deadline for a negotiation. For instance, agents can keep track of the food consumption rate in a refrigerator, and set the negotiation deadline to the date that the food will be exhausted. In addition, agents know their owners' preferences so that the weights can be automatically determined rather than relying on user input. For instance, if the agent finds that most of the electric appliances of its owner are of brand A, it is plausible to assume that brand A is its owner's favorite. When purchasing other appliances, the agent would give higher priority to that brand. While it is impossible for an agent to know everything about its owner, its ability to anticipate can substantially reduce the required user input. A user's role thus becomes to confirm, rather than to initiate or to specify, a task.

We propose a framework of automated negotiation systems on the basis of genetic algorithm and agent technology. Our proposed system can be used in conjunction with the electronic marketplace such as [1]. Fig. 1 illustrates the details of the idea.

In our proposed framework, agents proactively predict their owner's needs as well as their requirements. This feature is particularly important since need identification is often regarded as the most important stage in the consumer buying behavior (CBB) model [5]. Nevertheless, currently only very primitive event-alerting tools are available at some commercial web sites (e.g., Amazon.com will send an e-mail to its customer if there is any new book that may be of interest to them), and little research effort has been paid to this issue. The advent of agent technology now brings hope in addressing this problem. Alternatively, users may manually initiate a negotiation process. In either case, the agent fills out the default specification according to its owner's profile. It should note that users may modify these specifications and the agent also learns from the feedback.

Section snippets

GA (Genetic agent) model

Our basic GA negotiation model is based on the one proposed by [7]. Initially, a negotiating agent identifies itself as either a buyer or a seller and receives an offer from its opponent. If the offer is not satisfactory, the agent generates a counter-offer by the following procedure. The agent first sets up a population of chromosomes, each of which represents a candidate offer for the current round of negotiation. (In our experiments, this population contains 90 randomly generated chromosomes

Three concession-matching tactics

We study three types of commonly used concession-matching tactics suggested by [7]; namely, reciprocal tactic, exploiting tactic and cooperative tactic. Reciprocal tactics is the negotiation strategy that imitates the opponent's concessionary behavior. In other words, if the opponent concedes more, the agent concedes more too. In the exploiting tactic, an agent concedes less when its opponent is cooperative. In the cooperative tactic, an agent concedes in order to reach a consensus quickly.

From

Empirical results

In our experiments, we consider four different issues for negotiation: price, quantity, processing days and features. In the first experiment, we first investigate the effects of combining different concession-matching strategies on the average negotiation time and the obtained payoff. We set all four issues as “quite mind” for the testing purpose. As suggested from Table 1, the exploit strategy in general performs the best in terms of the payoff but takes up the most negotiation time. This

Conclusion

Automated negotiation has become increasingly important since the advent of electronic commerce. In this paper, we propose a genetic-agent-based automated negotiation system for electronic business. Unlike other negotiation systems, our proposed system is able to proactively anticipate the user needs and initiate a purchase process. According to the user's profile and schedule, the agent suggests a set of default parameters and asks its owner for confirmation. This approach thus minimizes the

Samuel Choi received both his Bachelor and Master degrees in computer science from University of Manitoba (Canada), and his Ph.D. degree in computer science from Hong Kong University of Science and Technology. He taught at Hong Kong University of Science and Technology and is currently a post-doctoral teaching fellow in the department of Computer Science at Hong Kong Baptist University. His primary research interests are electronic commerce, intelligent agents and machine learning.

References (19)

  • S.P.M. Choi, J. Liu, A dynamic mechanism for time-constrained trading, in: Proceedings of Fifth International...
  • B. Crabtree, N.R. Jennings, (Eds.), The Practical Application of Intelligent Agents and Multi-Agent Technology, London,...
  • K. DeJong

    Adaptive systems design: a genetic approach

    IEEE Trans. Syst., Man Cybernetics

    (1980)
  • D.E. Goldberg

    Genetic Algorithms in Search, Optimization, and Machine Learning

    (1989)
  • R.H. Guttman et al.

    Agent-mediated electronic commerce: a survey

    Knowledge Eng. Rev. J.

    (1998)
  • M.N. Huhns et al.

    Readings in Agents

    (1998)
  • R. Krovi, A.C. Graesser, Agent behaviors in virtual negotiation environments, IEEE Trans. Syst., Man, Cybernetics (Part...
  • R.J. Lewicki et al.

    Negotiation. Readings, Exercises, and Cases

    (1985)
  • T. Mitchell

    Machine Learning

    (1997)
There are more references available in the full text version of this article.

Cited by (0)

Samuel Choi received both his Bachelor and Master degrees in computer science from University of Manitoba (Canada), and his Ph.D. degree in computer science from Hong Kong University of Science and Technology. He taught at Hong Kong University of Science and Technology and is currently a post-doctoral teaching fellow in the department of Computer Science at Hong Kong Baptist University. His primary research interests are electronic commerce, intelligent agents and machine learning.

Jiming Liu is an Associate Professor of Computer Science at Hong Kong Baptist University (HKBU), a Senior Member of the IEEE, and a Member of the ACM. His areas of expertise are artificial intelligence, multi-agent systems, learning, adaptation and artificial life in software and systems, and autonomy-oriented computation (AOC). Dr. Liu earned the Master-of-Arts degree in Educational Technology from Concordia University, and the Master-of-Engineering and the Doctor-of-Philosophy degrees both in Electrical Engineering from McGill University in Montreal, Canada. He worked for some time as Software Engineer, Research Associate, and Senior Research Agent at R&D firms and government labs in Canada before joining the Computer Science Department of HKBU in 1993. In 1999, he was a Visiting Scholar in the Computer Science Department at Stanford University.

Ricky Chan obtained the Bachelor of Science (Hons.) in Computer Science (Computer System) from Hong Kong Baptist University (HKBU) in 1997. He is currently Internet Application Programmer in Interactive Communication Online Networks, Limited (ICON), mainly working on designing and implementing portal tools with servlets and JSP. His research interest is artificial intelligence, especially in machine learning.

View full text