for Retail Electronic Commerce
Robert H. Guttman and Pattie Maes
MIT Media Laboratory20 Ames Street, E15-301Cambridge, MA 02139{guttman,pattie}@media.mit.eduhttp://ecommerce.media.mit.edu
Abstractdepicted in Figure 1(a). There are an increasing number
Software agents help automate a variety of tasks of software agent tools available to merchants for including those involved in buying and selling products enhancing and differentiating their product offerings over the Internet. Although shopping agents provide online such as Firefly Network’s recommendation convenience for consumers and yield more efficient system [3, 8] and PersonaLogic’s buying guides [4]. markets, today’s first-generation shopping agents are These tools help consumers make buying decisions limited to comparing merchant offerings only on price within a specific merchant’s site. However, consumers instead of their full range of value. As such, they do a also compare product offerings across merchant disservice to both consumers and retailers by hiding boundaries as depicted in Figure 1(b). Due to the lower important merchant value-added services from transaction costs of online marketplaces and with the consumer consideration. Likewise, the increasingly
help of software shopping agents, consumers can easily
popular online auctions pit sellers against buyers in
distributive negotiation tug-of-wars over price. This perform cross-merchant product comparisons (whether paper analyzes these approaches from economic, merchants want this or not). behavioral, and software agent perspectives then proposes integrative negotiation as a more suitable
m1m2...approach to retail electronic commerce. Finally, we
identify promising techniques (e.g., multi-attribute utility theory, distributed constraint satisfaction, and p1conjoint analysis) for implementing agent-mediated integrative negotiation.
p21. Introduction
Online marketplaces are both an opportunity and a threat to retail merchants. They are an opportunity because they offer traditional merchants an additional channel to advertise and sell products to consumers thus potentially increasing sales. Forrester Research estimates that online retail sales were at about $600 million USD in 1996, will exceed $2 billion USD in 1997, and will reach $17 billion USD by 2001 [1]. In addition, online markets are more efficient than their physical-world counterparts thus lowering transaction costs for both merchants and consumers. For example, low transaction costs is one reason why Amazon.com [2], a virtual bookstore, can offer a greater selection and lower prices than its physical-world competitors.1.1.Cross-Merchant Product ComparisonsAs in the physical world, an online merchant prefers to have consumers shop only at its own Web site as
1
...(a) Within-Merchant Product Comparisons
m1p1p2...m2...(b) Cross-Merchant Product ComparisonsFigure 1 - Merchants prefer that consumers shop for products only within their own store (a). However, on the Internet, software shopping agents make it very easy for consumers to cross merchant boundaries and perform cross-merchant product comparisons (b).
1.2.Value-Added, Merchant Differentiation
and Market Power
m1m2...Although cross-merchant product comparisons are a
p1threat to merchant profitability, they are characteristic of
the retail marketplace and are here to stay. Knowing this, retailers add value to manufacturers’ products to p2distinguish themselves from their competitors. These value-added services include extended warranties, ...forgiving return policies, wide product selections, brand reputation, extensive service contracts, special gift Figure 2 - First-generation Cross-Merchant services, high product availability, superior customer Shopping Agentsservice and support, diverse payment, loan and leasing options, fast delivery times with low costs, promotions
and coupons, cross-manufacturer product configurations, problem warrants attention. Although free markets are etc. Depending on the product, these value-added inherently “nature red in tooth and claw” [6], this need services can be critical to a consumer’s buying decision not be the relationship between retailers and their
customers. Rather, we propose that a more cooperative regardless of the manner of shopping.
and personalized integrative negotiation approach
Merchant differentiation through added value is differentiates retailer’s offerings in online markets better necessary for merchants to exercise market power, the than today’s limited price-comparison shopping agents ability of a merchant to raise the price of a product above and unnecessarily hostile distributive negotiation (e.g., its marginal cost. In a fully competitive market, no one auction) approaches.has market power forcing prices down to the cost of producing the most expensive (marginal) unit [5].
Therefore, without merchant differentiation, retailers (and 1.3.Consumer Buying Behavior Modelother intermediaries) are reduced to competing on Consumer Buying Behavior (CBB) marketing marginal costs leaving little room for profit.research builds descriptive theories and models for
Unfortunately for online retailers, all of today’s analyzing consumers’ actions and decisions involved in first-generation cross-merchant shopping agents are buying and using goods and services. Guttman et al limited to comparing merchant offerings only on price augment traditional CBB research with concepts from
Software Agents research to accommodate electronic
instead of their full range of value as depicted in Figure
markets [7]. Table 1 lists all six stages of this CBB
2. This makes it hard (if not impossible) for merchants
model and gives representative examples of agent
to effectively differentiate themselves. This results in
systems that fall within this space.
inappropriately competitive retail markets and forces
Briefly, the Product Brokering stage comprises the merchants to compete almost entirely on marginal costs.
retrieval of information to help determine what to buy.
This paper suggests a reversal of this problematic
This encompasses the evaluation of product alternatives
trend in cross-merchant shopping agent approaches in
based on consumer-provided criteria. The result of this
order to restore merchant differentiation and thus their
stage is the “consideration set” of products. The
market power. With so much money at stake, this
$$Persona FireflyLogic1. Need Identification2. Product Brokering3. Merchant Brokering4. Negotiation5. Purchase and Delivery6. Product Service and Evaluation√√Bargain JangoFinderKasbahAuction Auction BotWeb√√√√√√√Table 1 - The six stages of the CBB model with representative examples of agent mediators [7].
2
Merchant Brokering stage combines this “consideration Consulting’s BargainFinder [9] was the first merchant set” with merchant-specific information to help brokering shopping agent. Given a specific music CD, determine who to buy from. This includes the evaluation BargainFinder requests its price (including shipping) of merchant alternatives based on consumer-provided from each of nine different online music catalogs using criteria (e.g., price, warranty, availability, delivery time, the same requests as a web browser. BargainFinder then reputation, etc.). The Negotiation stage is about how to presents its results to the consumer. As would be determine the terms of the transaction. In traditional expected from the discussion in section 1, several of the retail markets, price and other aspects of the transaction merchants preferred not to participate and blocked all are often fixed leaving no room for negotiation. In other price requests from BargainFinder as shown in Figure 3.markets (e.g., stocks, automobile, fine art, local It’s also interesting to note that CDLand initially markets, etc.), the negotiation of price or other aspects blocked BargainFinder agents but eventually decided to of the deal are integral to product and merchant compete on price. However, a visit to their site brokering.indicates that it has been deactivated for the past seven As noted in [7], this analysis of retail electronic months due to “new management” and “some initial
transition difficulties” [10]. We can only presume that a commerce represents an approximation and
simplification of complex behaviors. CBB stages often lack of merchant differentiation and market power lead to overlap and migration from one to another is sometimes their demise.nonlinear and iterative.However, is merchant differentiation still relevant for commodity-like and low-price markets such as music
CDs? Although there may be less of a need, properly
2.Price-Only Shopping Agents andpresented merchant differentiation can help consumers
Distributive Negotiation Agentsmake more educated buying decisions even in these There are several types of software shopping agents markets. For instance, when buying music CDs,
consumers may still want to consider product that assist consumers in making buying decisions.
Table 1 gives representative examples of agents that play availability, delivery times and costs, gift services, different roles in mediating online transactions. See [7] return policies, customer service, as well as promotions for a treatment of agent systems playing in the Product and coupons.Brokering stage of the CBB model.Excite’s Jango [11, 12] is similar to BargainFinder but with more product features to search across and more
shopping categories. The following sidebar describes the
After Product Brokering comes Merchant Brokering limitations of using Excite’s Jango Shopping Agent to in the CBB model shown in Table 1. Andersen buy a specific notebook computer.2.1.Price-comparison Shopping Agents
Figure 3 - BargainFinder requests prices of a given music CD from nine separate merchants and displays them to the consumer for a price comparison. However, three of the nine merchant sites are blocking BargainFinder’s price requests.
3
An Experience with Excite’s Jango Shopping AgentAbove are the results from Excite’s Jango of a search for an Apple PowerBook 3400C/200 notebook computer with 16 MB of RAM, a 2.0 GB hard drive, and a 16x CD-ROM drive. Assuming I’m considering buying this computer (and know what these features mean), who do I buy it from? According to Excite’s Jango, I can either buy the product from CDW for $3579.88 or from MicroWarehouse for $3799.00. All other things being equal (as Jango would have us believe), the rational decision is to buy the product from CDW for $219 less than from MicroWarehouse. But are all other things equal?After a while of “manual” investigation, I discover that Apple is having a promotion. If I buy this product from MicroWarehouse within the next four weeks, I’ll get a free 32MB RAM chip or a free Apple QuickTake Camera! It’s not clear whether CDW honors this promotion. Such a differentiation makes the merchant offerings more comparable. Even if CDW also honors the promotion, it would have been useful if Jango informed me of it — perhaps enticing me to buy the product when I may not have otherwise.More importantly, I also discover during my investigation that both merchants offer a 30 day return policy. However, if I’m unhappy with the product and return it to CDW, I’ll be charged a 15% restocking fee. MicroWarehouse doesn’t have a restocking fee. The $219 savings from buying it from CDW instead of from MicroWarehouse would have resulted in an extra $537 expense. It would have been useful if Jango allowed me to consider this information in my buying decision.I’m still considering a purchase, but what are the reputations of these merchants? Do they offer extended warranties, service contracts, loan options, or gift services? Is the product even available? If so, how fast can it be delivered? How much will that cost? What other goods and services do I need to configure the product appropriately for my needs? A good sales agent would answer these questions to assist me in making a more educated buying decision and offer more products and options for consideration. Jango is not assisting me in considering any merchant value add in this buying decision.2.2.Distributive Negotiation
Like the term “agent”, there is no consensus on the definition of the term “negotiation.” Economists, game theorists, business managers, political scientists, and artificial intelligence researchers each provide unique perspectives on its meaning. The business negotiation literature defines two types of negotiation: distributive The benefit of dynamically negotiating a price for a negotiation and integrative negotiation [13]. product instead of fixing it is that it relieves the seller Distributive negotiation is the decision-making process from needing to determine the value of the good a priori. of resolving a conflict involving two or more parties Rather, this burden is pushed into the marketplace itself. over a single mutually exclusive goal. The economics A resulting benefit of this is that limited resources are literature describes this more specifically as the effects allocated fairly — i.e., to those buyers who value them
4
on market price of a limited resource given its supply and demand among self-interested parties [5]. The game theory literature describes this situation as a zero-sum game where as the value along a single dimension shifts in either direction, one side is better off and the other is worse off [14].
most. As such, distributive negotiation mechanisms are common in a variety of markets including stock markets (e.g., NYSE and NASDAQ), fine art auction houses (e.g., Sotheby’s and Christie’s), flower auctions (e.g., Aalsmeer, Holland), and various ad-hoc haggling (e.g., automobile dealerships and commission-based electronics stores). More recently, software agents have been taught distributive negotiation skills (e.g., auctioneering and auction bidding skills) to help automate the Negotiation CBB stage of consumer-to-consumer and retail shopping over the Internet.
frugal – corresponding to a linear, quadratic, or exponential function respectively for increasing its bid for a product over time. The simplicity of these negotiation heuristics makes it intuitive for users to understand what their agents are doing in the marketplace.1 This was important for user acceptance as observed in a recent Media Lab experiment [15]. A larger Kasbah experiment is now underway at MIT allowing students to transact books and music [16].
AuctionBot [19, 20] is a general purpose Internet auction server at the University of Michigan.
Kasbah [15, 16] is a Web-based multi-agent AuctionBot users create new auctions to buy or sell classified ad system where users create buying agents and products by choosing from a selection of auction types selling agents to help transact goods. These agents and specifying its parameters (e.g., clearing times, automate much of the Merchant Brokering and method for resolving bidding ties, the number of sellers Negotiation CBB stages for both buyers and sellers. A permitted, etc.) as shown in Figure 4. Buyers and sellers user wanting to buy or sell a good creates an agent, can then bid according to the multilateral distributive gives it some strategic direction, and sends it off into a negotiation protocols of the created auction. In a centralized agent marketplace. Kasbah agents proactively typical scenario, a seller would bid a reservation price seek out potential buyers or sellers and negotiate with after creating an auction and let AuctionBot manage and them on behalf of their owners. Each agent’s goal is to enforce buyer bidding according to the auction protocols complete an acceptable deal, subject to a set of user-and parameters.specified constraints such as a desired price, a highest (or AuctionBot also provides an application lowest) acceptable price, and a date by which to complete programmable interface (API) for users to create their the transaction.own software agents to autonomously compete as buyers Negotiation between buying and selling agents in or sellers in the AuctionBot marketplace. This API Kasbah is bilateral, distributive, and straightforward. permits AuctionBot to enforce auction protocols and After buying agents and selling agents are matched, the provides a semantically sound communication interface only valid action in the distributive negotiation protocol
Unlike other multi-agent marketplaces [18], Kasbah does is for buying agents to offer a bid to sellers. Selling
agents respond with either a binding “yes” or “no”. not concern itself with optimal strategies or convergence
properties. Rather, Kasbah provides more descriptive
Given this protocol, Kasbah provides buyers with one of
strategies that model typical haggling behavior found in
three negotiation “strategies”: anxious, cool-headed, and classified ad markets.
1
5
to the marketplace. However, as with the similar Fishmarket system from the Artificial Intelligence Research Institute in Barcelona [21, 22], it is left to the buyers and sellers to encode their own bidding strategies.2
Agent systems like Kasbah and AuctionBot are useful for building prescriptive theories for coordination among heterogeneous agents with (partially) predictable system-wide dynamics. However, as described next, distributive negotiation auctions are not well-suited for retail markets.2.3.Auction Fever
Two of the original (non-academic) auction Web sites are OnSale [25] and eBay’s AuctionWeb [26] and are still very popular. Likely reasons for their popularly include their novelty and entertainment value in negotiating the price of everyday goods, as well as the potential of getting a great deal on a wanted product. In any case, the popularity of OnSale and eBay’s AuctionWeb has quickly spawned an already competitive and growing industry. Whereas once auctions were in themselves novel merchant differentiators, with the rapid proliferation of online auctions, this differentiation has waned. Yahoo! lists more than 90 active online auctions today [27]. Forrester Research reports that auctions will be core to making business-to-business transactions more dynamic, open and efficient [28]. online auctions like FastParts [29] and FairMarket [30] are already making this happen in the semiconductor and computer industries.
What’s most relevant here is that many online auctions are augmentations to retail sites with retailers playing the roles of both auctioneer and seller (i.e., a sales agent). For example, First Auction [31] is a service of Internet Shopping Network, one of the first online retailers. Cendant’s membership-driven retail site, netMarket [32], has also recently added auctions to its repertoire of online services. New auction intermediaries such as Z Auction [33] offer their auction services to multiple manufactures and resellers as a new sales channel.
bid, determining the optimal bidding strategy is non-trivial3 and, more importantly, can be financially adverse. In fact, in first-price open-cry auctions (i.e., highest bid wins the good for that price), the winning bid is always greater than the product’s market valuation. This is commonly known as “winner’s curse” as depicted in Figure 5. This problem is exacerbated in retail auctions where buyers’ valuations are largely private4. Buyers with private valuations tend to (irrationally) skew bids even further above the product’s true value.
marketvaluation
winningbid# of bidspricewinner’scurse
Figure 5 - “Winner’s curse” is the paradox that the winning bid in an auction is greater than the product’s market valuation. This occurs in all first-price, open-cry auctions – the most prevalent type on the Internet.
With this much “auction fever,” you would think
Another customer dissatisfaction problem owing to that auctions are a panacea for retail shopping and
selling. On the contrary, upon closer look we see that online auctions is the long delay between the start of the auctions have rather hostile characteristics. For
Factors to be considered include information asymmetry, example, although the protocols for the two most
risk aversion, motivation and valuation.
prevalent types of online auctions, first-price open-cry
The motivation of a buyer with private-valuation is to
English and Yankee [34], are simple to understand and acquire goods for personal consumption (or for gifts).
34
2
Although winner’s curse is a short-term financial
benefit to retailers, it can be a long-term detriment due to the eventual customer dissatisfaction of paying more than the value of a product. Two universal auction rules that compound this problem are: (1) bids are non-retractable and, worse yet, (2) products are non-returnable. This means that customers could get stuck with products that they’re unhappy with and paid too much for. In short, online auctions are less lucrative and far less forgiving than would be expected in retail shopping.
Although not currently deployed as a real-world shopping system, Fishmarket has hosted tournaments to compare opponents’ hand-crafted bidding strategies [23] along the lines of Axelrod’s prisoner’s dilemma tournaments [24].This is in contrast to a buyer with common-valuation (e.g., in stock) where the motivation is to make money through the buying and later reselling of goods which have no other intrinsic value to the buyer.
6
Negotiation CBB stage and the end of the Purchase and the value of their goods a priori.Delivery CBB stage. For example, due to In addition, where sellers may have shills, buyers communication latency issues and wanting a critical may collude by forming coalitions. A buyer coalition mass of bidders, the English and Yankee auction is a group of buyers who agree not to outbid one protocols as implemented over the Internet extend over another. In a discriminatory (i.e., multi-good) auction, several days. This means that after a customer starts the result of this is that the coalition can buy goods for bidding on a product, she/he must continuously bid for less than if they competed against one another thus the product (or have a shopping agent do it as provided unfairly cheating the seller. The coalition can then by AuctionWeb) up until the auction closes several days distribute the spoils amongst themselves (e.g., evenly, later. This does not cater to impatient or time-by holding a second private auction, etc.). As with constrained consumers.5 To make matters worse, only shills, collusion through buyer coalitions is also the highest bidder(s) of an auction can purchase the considered illegal. However, as with shills, it can be auctioned good meaning that the other customers need to hard to detect buyer collusion, especially in online wait until the good is auctioned again and then restart the markets where bidders are virtual. In fact, Multi-Agent Negotiation CBB stage.6 Additionally, since bids are Systems research has developed technologies that can non-retractable and binding, consumers are unable to efficiently form coalitions even among previously reconsider earlier brokering decisions during this delayed unknown parties [36] — posing an additional threat to negotiation stage.online retail auctions.There are other buyer concerns with English and Yankee style auctions such as shills. Shills are bidders who are planted by sellers to unfairly manipulate the market valuation of the auctioned good by raising the bid to stimulate the market. Although deemed illegal in all auctions, shills can be hard to detect especially in the virtual world where it is relatively inexpensive to create virtual identities (and thus virtual shills). Also, there is usually no negative consequence to the seller if one of his/her shills (accidentally) wins the auction.
As explained, online auctions are unnecessarily hostile to customers and offer no long-term benefits to merchants. Essentially, they pit merchant against customer in price tug-of-wars. This is not the type of relationship merchants prefer to have with their customers [38]. Unlike most consumer-to-consumer and commodity markets, merchants often care less about profit on any given transaction and care more about long-term profitability. This ties directly to customer satisfaction and long-term customer relationships. The more satisfied the customer and intimate the customer-merchant relationship, the greater the opportunity for repeat customer purchases and additional purchases through direct referrals and indirectly through positive reputation.
Distributive negotiation auctions in retail markets also pose problems for merchants. Although auctions can relieve merchants of the burden of establishing prices for limited resources (e.g., fine art and stocks), this benefit is less realizable for production goods as in retail markets. Unlike fine art, for example, it is relatively And as with price-only shopping agents, distributive easy to determine the marginal costs of production negotiation auctions focus the consumers’ attention goods.7 If auctioning these goods, however, it is non-solely on a product’s price rather than its full range of trivial for the merchant to determine the optimal size of value. This is a disservice to both consumers and the auctioned lots and the frequency of their auction [35]. merchants because, as with price-comparison shopping Such a determination requires an understanding of the agents, it hides important merchant added value from demand for the good since it directly affects inventory consumers’ consideration.9 Also, by only negotiating management and indirectly affects production schedule.8 over price, merchants lose an opportunity to differentiate Therefore, retailers are still burdened with determining themselves during the earlier Merchant Brokering and
56
7
8
In fact, such delays are the antithesis to impulse buying.Even in traditional static catalog retail (as well as Continuous Double Auctions), consumers can purchase products immediately.
Granted, the pricing of retail products can get involved. This is where marketing tactics come into play such as branding, market segmentation, price discrimination, etc.
This relates directly to the just-in-time (JIT) concept for manufacturing, inventory, and retailing [37]. However, it is not yet clear how best to gauge demand in JIT (e.g., through negotiation or sales).
Product Brokering CBB stages. Ultimately, by shortsightedly succumbing to “auction fever,” retail merchants may be instrumental in bringing about their own demise. By promoting auctions as appropriate retail negotiation mechanisms, it strips themselves of differentiation and exposes their markets to greater
9
For example, Gerry Heller, CEO of FastParts - an online auction for semiconductors, was quoted in a recent Forrester Research report as admitting that even in this commodity-like market “availability is more important than price” when it comes to auctioning semiconductors.
7
competition thus nullifying their market power and which is a win-lose type of negotiation as discussed in profit.section 2.2. Also as discussed, all auctions are forms of
distributive negotiation and are therefore win-lose types
3.Integrative Negotiation Agentsof negotiation.There is a tremendous amount of literature on how Desired retail merchant-customer relationships and to sell retail products. No one approach is correct as it interactions can be described in terms of integrative depends upon a number of factors including the type of negotiation — the cooperative process of resolving product and demographics of its intended audience. multiple interdependent, but non-mutually exclusive Likewise, there is no one correct way to shop. People goals. A merchant’s primary goals are long-term have different goals, knowledge, preferences, constraints, profitability through selling as many products as influences, and attitudes during any given shopping possible to as many customers as possible for as much experience. One type of shopping is cross-merchant money as possible with as low transaction costs as product comparisons (see section 1.1). It assumes a possible. A customer’s primary goals are to have their (partially) rational shopper who is concerned with personal needs satisfied through the purchase of well-buying the merchant offering that best meets his/her suited products from appropriate merchants for as little needs given an invested amount of time and effort.money and hassle (i.e., transaction costs) as possible.
Cross-merchant product comparisons are conducive An integrative negotiation through the space of merchant to software agent mediation by assisting the shopper in offerings can help maximize both of these sets of goals. any of the Product Brokering, Merchant Brokering, and From a merchant’s perspective, integrative negotiation is Negotiation stages of the Consumer Buying Behavior about tailoring its offerings to each customer’s model. However, some agent-mediation approaches are individual needs resulting in greater customer better than others. We argue in section 2.1 for shopping satisfaction. From a customer’s perspective, integrative agents that can perform value-comparisons, not just negotiation is about conversing with retailers to help price-comparisons. In section 2.3, we argue for sales compare merchant offerings across their full range of agents that can negotiate over the full range of a value resulting in mutually rewarding and hassle-free
shopping experiences.merchant’s added value rather than just price.
We propose an integrative negotiation approach to 3.2.Multi-Objective Decision Analysis
cross-merchant product comparisons. This approach and Multi-Attribute Utility Theorypromotes negotiation between consumer-owned
Multi-objective decision analysis prescribes theories shopping agents and merchant-owned sales agents across
each product’s full range of value. The rest of this for quantitatively analyzing important decisions section discusses integrative negotiation and identifies involving multiple, interdependent objectives from the
perspective of a single decision-maker [40]. This promising techniques for its implementation.
analysis involves two distinctive features: an uncertainty analysis and a utility (i.e., preference) analysis.
3.1.Integrative NegotiationTechniques such as bayesian network modeling aid
As introduced in section 2.2, the business uncertainty analysis. Multi-attribute utility theory negotiation literature defines two types of negotiation: (MAUT) analyzes preferences with multiple attributes.distributive negotiation and integrative negotiation. Examples of uncertainty in retail shopping are “will Integrative negotiation is the decision-making process she like this product as a gift?” and “how much do I trust of resolving a conflict involving two or more parties this merchant?” Such uncertainties weighed against over multiple interdependent, but non-mutually other factors play a part in consumers’ buying decisions. exclusive goals [13]. The study of how to analyze From a merchant’s perspective, analyzing an uncertainty multi-objective decisions comes from economics like “what will be the demand for this product?” is vital research and is called multi-attribute utility theory for pricing products and managing inventory.(MAUT) [40]. The game theory literature describes
Often, decisions have multiple attributes that need integrative negotiation as a non-zero-sum game where as
the values along multiple dimensions shift in different to be considered. For example, in retail shopping, the directions, it is possible for all parties to be better off price of a product could be important, but so could its
delivery time. What is the relationship and tradeoff [14].
between these two? Figure 6 gives a simple example of
In essence, integrative negotiation is a win-win type this.of negotiation. An example of this is depicted in Figure
Multi-objective decision analysis and MAUT can 7. This is in stark contrast to distributive negotiation 8
total price(and have) been used to tackle many different types of quality, airport location, heroin addiction treatment, decision problems including electrical power vs. air medical diagnostic and treatment, business problems,
political problems, etc. These theories have also been instantiated in computer systems. The PERSUADER consumer’s utility
system at Carnegie Mellon University, for example, merchant’s utility
integrates Case-Base Reasoning and MAUT to resolve conflicts through negotiation in group problem solving
overnightsettings [39]. Logical Decisions for Windows (LDW)
2-3 daysby Logical Decisions, Inc. [41] is a general-purpose
decision analysis tool for helping people think about and 7-10 daysanalyze their problems. Figure 7 shows LDW at work on a retail purchase decision problem.
matchdelivery timeFigure 6 - This graph plots a consumer’s and a merchant’s multi-attribute utilities for a product’s total price vs. delivery time (in days). In this example, the merchant offers three delivery options at different price points of which the “2-3 days” option best matches the consumer’s utility profile.
LDW falls within the Product Brokering stage of our CBB model. However, MAUT tools such as LDW can also be applied to the Merchant Brokering CBB stage by formulating a new problem to analyze merchant value add for the winning product (i.e., considered set) of the Product Brokering stage. If certain pragmatic issues concerning MAUT’s appropriateness for real-time Internet-based bilateral negotiations can be allayed, then MAUT techniques are contenders for decision support in agent-mediated integrative negotiation strategies for online retail markets.
Figure 7 - A screenshot of Logical Decisions for Windows (LDW). This screenshot shows the results of a computer purchase decision after LDW captured the decision-maker’s utilities across multiple product attributes. One results window shows the product rankings and the other a side-by-side comparison of two product contenders.
9
Figure 8 - Screenshots of PersonaLogic assisting a customer select automobile features with results.
3.3.Distributed Constraint SatisfactionMAUT analyzes decision problems quantitatively through utilities. Constraint Satisfaction Problems (CSPs) analyze decision problems more qualitatively through constraints. A CSP is formulated in terms of variables, domains, and constraints. Once a decision problem is formulated in this way, a number of general purpose (and powerful) CSP techniques can analyze the problem and find a solution [42].
Finite-domain CSPs are one type of CSP and are composed of three main parts: a finite set of variables, each of which is associated with a finite domain, and a set of constraints that define relationships among variables and restricts the values that the variables can simultaneously take. The task of a CSP engine is to assign a value to each variable while satisfying all of the constraints. A variation of these “hard” constraints is the ability to also define “soft” constraints (of varied importance) which need not be satisfied. The number, scope, and nature of the CSP’s variables, domains, and constraints will determine how constrained the problem is and, for a given CSP engine, how quickly a solution (if any) will be found.
problems. In retail markets, CSP techniques can be used to encode hard constraints such as “I’m not willing to spend more than $2,000 for this product,” and soft constraints such as “availability is more important to me than price.” Even constraints such as “I prefer the Gateway 2000 P5-90 over the Dell Dimension XPS P (but I don’t know why)” are legitimate. PersonaLogic (see Figure 8) uses CSP techniques to help shoppers evaluate product alternatives in the Product Brokering stage of our CBB model. Given a set of constraints on product features, PersonaLogic filters products that don’t meet the given hard constraints and prioritizes the remaining products using the given soft constraints. This approach can also be applied to sales configuration systems such as Dell’s “Build Your Own System” [43] and Trilogy’s Selling Chain™ [44].
An important side-benefit of CSPs is that they can clearly explain why they made certain decisions such as removing a product from the results list (e.g., “Product X is not an option because it has only 16 MB of RAM and you specified that the product should have at least 32 MB of RAM.”). This feature is important because it relates to consumer trust. Trust is partially achieved by the shopping agent exhibiting somewhat predictable
Many problems can be formulated as a CSP such as behavior and being able to explain its decisions.scheduling, planning, configuration, and machine vision
10
As with LDW, PersonaLogic can likely be extended set” of products.12
into the Merchant Brokering stage of the CBB model. In However, in order to make a product selection, fact, it may even be possible to extend PersonaLogic consumers need to identify differences in product into the Negotiation stage by using Distributed attributes. It may be better for a user to just express Constraint Satisfaction Problem (DCSP) techniques. these attribute preferences rather than spend time making DCSPs are similar to CSPs except that variables and a series of product choices which will (at best) infer the constraints are distributed among two or more loosely-same preferences. Conjoint analysis also suffers from coupled agents [45]. This appears to map well to the not dealing well with noisy or inconsistent data (which retail case where consumers and merchants each have are very common in user surveys), not being conducive their respective set of constraints on merchant offerings.to changes in product preferences, and being time-However, DCSPs have been designed for fully consuming, redundant, and boring for the consumer. As cooperative group problem solving situations. such, although conjoint analysis is appropriate for Although integrative negotiations are far more identifying new product features and segmenting cooperative than distributive negotiations, DCSP markets, it appears less appropriate as the sole techniques may require more cooperation than is mechanism for extracting utility preferences for appropriate for merchant-customer interactions. For integrative negotiations in retail electronic markets.example, a customer may not be willing to divulge her There are numerous statistical, search, and heuristic reservation value (e.g., a willingness to pay up to approaches that can also learn preferences and patterns of $2,000 for a computer) to a merchant for fear of first-user behavior. In fact, a tenet of artificial intelligence degree price discrimination with the merchant (unfairly) (AI) is learning. Specific AI fields of inquiry include capturing all of the surplus in the market. However, inductive learning, genetic algorithms, classifier first-degree price discrimination is tenuous in markets systems, case-based reasoning, neural networks, and a with monopolistic competition — i.e., a market with a variety of other machine learning and adaptive behavior large number of firms selling similar but differentiated theories and technologies [47, 48].products with no significant barriers to entry — which characterizes most retail markets [5]. This suggests that
DCSP techniques may not be overly cooperative for 4.Conclusionbilateral integrative negotiations in retail markets.10This paper analyzed the state-of-the-art in agent-mediated retail electronic commerce. We first looked at
how price-only shopping agents are a disservice to both 3.4.Conjoint Analysis and
consumers and retailers by hiding important merchant Machine Learning
value add from consumer consideration. We then
Conjoint analysis is a popular marketing tool to explored how distributive negotiation techniques (e.g., help identify and market new product features [46]. The online auctions) are considerably more hostile to both approach involves repeatedly surveying respondents for consumers and merchants than would be expected in the preferred product given two or more product choices. retail markets (in spite of their increasing popularity). This is in contrast to rating products (e.g., in automated
Finally, we proposed a new integrative negotiation collaborative filtering) or specifying requirements on
approach to retail electronic commerce. We described product attributes (e.g., in constraint satisfaction).
Rather, respondents jointly consider11 and relatively rank how techniques such as multi-attribute utility theory, product choices. Conjoint analysis then infers which distributed constraint satisfaction, and conjoint analysis product attributes are most important to the consumer could be harnessed for allowing consumer’s to relieving the consumer of specifying these features integratively negotiate over a product’s full range of explicitly. Also, by being forced to make product value. From a merchant’s perspective, integrative decisions, consumers avoid unreasonable product negotiation is about tailoring its offerings to each attribute combinations — e.g., the most robust feature customer’s individual needs resulting in greater customer set and the lowest price. This is a benefit over CSPs satisfaction. From a customer’s perspective, integrative which allow consumers to specify unreasonable product negotiation is about conversing with retailers to help attribute combinations resulting in an empty “considered compare merchant offerings across their full range of
value resulting in mutually rewarding and hassle-free
Full cooperation does not necessitate full disclosure. For shopping experiences.
10
11
example, merchants need not divulge their profit margins. However, full cooperation does assume soundness of trust - i.e., false advertising isn’t permitted.Conjoint is a contraction of “consider jointly.”
12
However, there are CSP techniques to automatically relax constraints in over-constrained problems.
11
5. Acknowledgements
[19]P. Wurman, M. Wellman, and W. Walsh. “The
Michigan Internet AuctionBot: A Configurable
We would like to thank Alex Kleiner III, Fernanda Auction Server for Human and Software Agents.” To Viegas, Natalia Marmasse, and Alexandros Moukas for appear, Proceedings of the Second International their help with this paper.Conference on Autonomous Agents (Agents’98), May
1998.
6. References
[20]AuctionBot URL:
[1]Forrester Research Report. On-Line Internet Spending. [21]Fishmarket URL:
1997. Firefly Network URL: McGraw-Hill, Inc., 1996. A. L. Tennyson. “In Memoriam” (LVI, 15). R. Guttman, A. Moukas, and P. Maes. “Agent-mediated Electronic Commerce: A Survey.” To appear, Knowledge Engineering Review, June 1998.U. Shardanand and P. Maes. “Social Information Filtering: Algorithms for Automating 'Word of Mouth'.” Proceedings of the Computer-Human Interaction Conference (CHI’95), Denver, Colorado, May 1995. [22]J. Rodriquez, P. Noriega, C. Sierra, and J. Padget.. “FM96.5: A Java-based Electronic Auction House.” Proceedings of the Second International Conference on the Practical Application of Intelligent Agents and Multi-Agent Technology (PAAM’97). London, UK, April 1997.[23]J. Rodriguez, F. Martin, P. Noriega, P. Garcia, and C. Sierra. “Competitive Scenarios for Heterogeneous Trading Agents.” To appear in Proceedings of the Second International Conference on Autonomous Agents (Agents'98).[24]R. Axelrod. The Evolution of Cooperation. Harper Collins, 1984.[25]OnSale URL: [26]eBay’s AuctionWeb URL: Technology Strategies: Sizing Intercompany Commerce, vol. 1, no. 1. July 1997.[29]FastParts URL: Journal of Economic Perspectives, pp. 3-22. Summer 19.[35]C. Beam, A. Segev, and J. G. Shanthikumar. “Electronic Negotiation through Internet-based Auctions.” CITM Working Paper 96-WP-1019, December 1996.[36]T. Sandholm and V. Lesser. “Coalition Formation among Bounded Rational Agents.” 14th International Joint Conference on Artificial Intelligence (IJCAI’95), Montreal, Canada, 1995.[37]G. Morgenson. “The Fall of the Mall.” Forbes, May 24, 1993.[38]Forrester Research Report. “Affordable Intimacy Strengthens On-Line Stores.” September, 1997. [8] [9]BargainFinder URL: [12]R. Doorenbos, O. Etzioni, and D. Weld. “A Scalable Comparison-Shopping Agent for the World Wide Web.” Proceedings of the First International Conference on Autonomous Agents (Agents’97). Marina del Rey, CA, February 1997.[13]R. Lewicki, D. Saunders, and J. Minton. Essentials of Negotiation. Irwin, 1997.[14]J. Rosenschein and G. Zlotkin. Rules of Encounter: Designing Conventions for Automated Negotiation among Computers. MIT Press, 1994.[15]A. Chavez, D. Dreilinger, R. Guttman, and P. Maes. “A Real-Life Experiment in Creating an Agent Marketplace.” Proceedings of the Second International Conference on the Practical Application of Intelligent Agents and Multi-Agent Technology (PAAM’97). London, UK, April 1997.[16]Kasbah URL: Market: Institutions, Theories, and Evidence. Addison-Wesley, New York, 1993.[18]C. Sierra, P. Faratin, and N. Jennings. “A Service-Oriented Negotiation Model Between Autonomous Agents.” Proceedings of the Eighth European Workshop on Modeling Autonomous Agents in a Multi-Agent World (MAAMAW’97). Ronneby, Sweden, May 1997. 12 [39]K. Sycara. “The PERSUADER.” In The Encyclopedia of Artificial Intelligence. D. Shapiro (ed.), John Wiley and Sons, January, 1992.[40]R. Keeney and H. Raiffa. Decisions with Multiple Objectives: Preferences and Value Tradeoffs. John Wiley & Sons, 1976.[41]Logical Decisions URL: [42]E. Tsang. Foundations of Constraint Satisfaction. Academic Press, 1993.[43]Dell “Build Your Own System” URL: [44]Trilogy’s Selling Chain™ URL: [45]M. Yokoo and E. Durfee. “Distributed Constraint Satisfaction for Formalizing Distributed Problem Solving.” Proceedings of the 12th IEEE International Conference on Distributed Computing Systems, 1992.[46]Crane, M. “Conjoint Analysis: A Guide for Designing & Interpreting Conjoint Studies.” Austin Texas: IntelliQuest, Inc., 1991.[47]J. Carbonell (ed.). Machine Learning: Paradigms and Methods. MIT Press, 1990.[48]S. Russell and P. Norvig. Artificial Intelligence: A Modern Approach. Prentice Hall, 1995. 13 因篇幅问题不能全部显示，请点此查看更多更全内容
Copyright © 2019- zrrp.cn 版权所有 赣ICP备2024042808号-1
违法及侵权请联系：TEL:199 1889 7713 E-MAIL:2724546146@qq.com
本站由北京市万商天勤律师事务所王兴未律师提供法律服务