Agent mediated Electronic commerce: An MIT Media Laboratory ...

$JHQWPHGLDWHG(OHFWURQLF&RPPHUFH $Q0,70HGLD/DERUDWRU\3HUVSHFWLYH $OH[DQGURV0RXNDV5REHUW*XWWPDQ*LRUJRV=DFKDULDDQG3DWWLH0DHV 6RIWZDUH$JHQWV*URXS0,70HGLD/DERUDWRU\&DPEULGJH0$86$ 7HO )D[ (PDLOPRX[JXWWPDQO\VLSDWWLH#PHGLDPLWHGX 85/KWWSHFRPPHUFHPHGLDPLWHGX

Abstract Software agents, semi-intelligent autonomous tools, will play an increasing role in electronic commerce applications. In this paper we give an overview of the work conducted at MIT’s Media Laboratory on different types of agents for electronic commerce: from consumer to consumer “smart” classified ads systems, to merchant agents that provide integrative negotiation capabilities; from agents that facilitate expertise brokering to distributed reputation facilities; and from point-of-sale comparive shopping agents to mobile shopping systems.

Keywords: Electronic Commerce; Software Agents; Virtual Marketplaces; Merchant Differatiation; Reputation Mechanisms; Comparison Shopping; Knowledge Brokering

0RXNDV*XWWPDQ=DFKDULD0DHV,-(&6XEPLVVLRQ

,QWURGXFWLRQ

,QWURGXFWLRQ Over the past two years, a new kind of software application has appeared based on a synthesis of ideas from artificial intelligence, human-computer interaction and electronic transactions: agents that help mediate electronic commerce activities. Agents differ from “traditional” software in that they are personalized, autonomous, proactive, and adaptive. These qualities make agents particularly useful for the information-rich and process-rich environment of electronic commerce. Shopping activities require a large effort from a consumer and include searching for parties interested in selling or buying what the consumer wants to buy or sell (e.g., by sifting through catalogs, advertisements in newspapers and television, shelves in stores, etc.), comparing prices and other features of the good or service to help make an optimal purchase decision, and exchanging currency for a product through some agreed upon, and ideally secure, channels. Electronic commerce - in particular, the buying and selling of goods, financial vehicles, and services on the Internet - has so far fallen short of its potential of redefining the marketplace. Reasons for this include consumer buying habits, security and trust concerns, and uncertain market models. In addition to streamlining traditional transactions, agents enable new types of transactions. For example, the elusive one-to-one marketing becomes more of a reality when consumer agents capture and share (or sell) consumer demographics. Prices and other transaction dimensions need no longer to be fixed; selling agents can dynamically tailor merchant offerings to each consumer. Economies of scale become feasible in new markets when agents negotiate on special arbitration contracts. Dynamic business relationships will give rise to more competitively agile organizations. It is these new opportunities combined with substantial reduction in transaction costs that will revolutionize electronic commerce. This paper is organized as follows: section two discusses background and motivation issues and how they relate to our work. Section three introduces our first generation of agent marketplaces and discusses the results from a real world experiment we conducted. Section four outlines our current work in different aspects of agent-mediated


%DFNJURXQGDQG0RWLYDWLRQ

electronic commerce. Finally, section five contains our summary and concluding remarks.

%DFNJURXQGDQG0RWLYDWLRQ One of the first domains that web-based electronic commerce focused on was consumer to consumer applications that resembled classified ads. Some of the first sites allowed users to perform keyword searches on the ads [1], some had their ads nicely categorized, making it easier to find ones of interest [2], while other sites had more advanced searching capabilities. For example, Stanford’s Infomaster [3] allowed consumers to fill out a form that precisely describes the kind of apartment they are looking for. Infomaster searches the ad databases of several local newspapers and returns those which describe apartments that match the user's specification. These “classified ad” sites provide tools to help the consumer find ads of interest. Certainly, such tools are useful. Yet they only assist with one step in the multi-step process of buying and selling, that of finding ads which match what one is looking for. The idea behind our first system, Kasbah was to help users with the other major step in the process, the negotiations between buyer and seller, by providing agents which can autonomously negotiate and make the “best possible deal” on the consumer behalf. In Kasbah we try to address issues of trust and distributed reputation mechanisms in electronic commerce both from a buyer’s and from a seller’s perspective. Unlike consumer to consumer electronic commerce, the early stages of business to consumer (retail) electronic commerce were not problem-free. The majority of the merchants established an on-line presence very similar to their static mail-order catalogs. The product information was static, there was no connection to their back-office and legacy systems and their on-line store and the transaction had to be finalized and the payment be made through a human operator. That initial state of affairs had reached a more or less steady state, with consumers “browsing” through the merchants sites (in a way similar to traditional “window shopping”), and the merchants themselves competing with each other in terms of pricing, service, warranty, web site presentation and design, etc. However, the introduction of 0RXNDV*XWWPDQ=DFKDULD0DHV,-(&6XEPLVVLRQ

%DFNJURXQGDQG0RWLYDWLRQ

the first consumer shopping agents changed drastically this fragile modus operandi. Consumer Shopping Agents were able to collect prices from a large number of merchants with no effort from the part of the consumer in almost zero time; moreover, they were programmed to look just for the best price. They didn’t take into consideration any of the additional merchant offerings mentioned above. BargainFinder [4], one of the first agents for price comparison offered valuable insights into the issues involved in price comparison in the on-line world. For example, some of the on-line CD merchants accessed by BargainFinder blocked all of its price requests for network load reasons and because merchants inherently did not wanted to compete on price alone (later on, as the BargainFinder site was gaining momentum and users, merchants explicitly asked to be added in the BargainFinder system). Value added services that merchants offered on their web site were being bypassed by BargainFinder and therefore not considered in the consumer’s buying decision. This was one of the first lessons learned in the electronic commerce arena: merchants don’t want to be compared just in product price terms. A while later, a new type of consumer agents, like Jango [5], [26] hit the market which managed to “solve” the merchant blocking issue by having the product requests originate from each consumer’s web browser instead of from a central site as in BargainFinder. This way, requests to merchants from a Jango-augmented web browser appeared as requests from real consumers. This was another lesson learned: The merchants with a web presence will be forced to interoperate with agents. This kind of “aggressive interoperability” makes it convenient for consumers to shop for commodity products but does not leave merchants with many options. If merchants provide on-line catalogs, they will be accessed by agents whether merchants want them to or not. In order for merchants to survive in the on-line commerce environment they need to be able to differentiate themselves from other merchants. Our Tete-a-Tete system addresses the merchant differentiation issue and tries to give merchants a competitive edge by providing them with agents that will on-the-fly negotiate with consumer agents on a customized feature space that will include many more attributes than just price. However, this is not as easy as it sounds given the existing state of merchant on


.DVEDK7KH)LUVW*HQHUDWLRQRI$JHQW0DUNHWSODFHV

line stores, which are designed and optimized to be accessed by humans. The vast majority of web pages are currently written in HTML (hypertext markup language) which is a data format language. In contrast, XML (extensible markup language) is a data content meta-language allowing for the semantic tagging of data [6], [7] and [8]. Microsoft and Netscape include support for XML with style sheets in their next generation of web browsers to help augment HTML with XML tags to create the next generation of WWW on-line stores what are readable both by humans and agents. Finally, an aspect of electronic commerce where agent technology application are very promising is that of services. One of our systems focuses on agents that can act as brokers that facilitate knowledge transfer and just-in-time consulting types of transactions.

.DVEDK7KH)LUVW*HQHUDWLRQRI$JHQW0DUNHWSODFHV Since 1994 we have embarked upon a research program to create electronic agent marketplaces. Our first system, Kasbah, is an electronic marketplace where agents buy and sell to one another on behalf of consumers [9]. Kasbah is a consumer to consumer electronic commerce system and is based on continuous double auction mechanisms. In a Kasbah environment we think of a selling agent as being analogous to a classified ad. When a consumer creates a new selling agent, it gives a description of the item they want it to sell. Unlike the traditional classified ad, though, which sits passively in its medium and waits for someone to notice it, Kasbah’s selling agents are pro-active. Basically, they try to sell themselves, by going into the marketplace, contacting interested parties (namely, buying agents) and negotiating with them to find the best deal. A selling agent is autonomous in that, once released into the marketplace, it negotiates and makes decisions on its own, without requiring consumer intervention. The consumer does have high-level control of its behavior. When the consumer creates a new selling agent, they set several parameters to guide it as it tries to sell the specified item. These parameters are:



Desired date to sell the item by. People usually have a deadline by which they want to sell something. For example, a graduating student might want to sell their bicycle before they leave school, because they cannot take it with them. Desired price. This is the price the consumer would like to sell their good for. Lowest acceptable price. This is the lowest price the consumer will sell their good for. If the consumer has junk in their basement that they want to get rid of, they may set the desired price rather high, hoping someone might be willing to pay it, and also set the lowest acceptable price to a more realistic level. On the other hand, a person willing to accept nothing less than their asking price would set the lowest acceptable price to be the desired price.

)LJXUH

%DVLF$JHQWV6WUDWHJLHV

These parameters define the agent’s goal: to sell the item in question for the highest possible price --- ideally, the desired price, but as low as the lowest acceptable price, if that is what it takes to attract buyer interest in the time frame given. Exactly how to achieve this goal is left to the agent.The appropriate metaphor here is that of a personal assistant [11]. You tell your personal assistant what you would like to be done (“sell this for the best possible price''), and trust it to figure out how to accomplish this task, free-



ing your time and energy for more interesting pursuits. In addition, we hope that agents might be able to sell (and buy) goods better (e.g. at a higher price) than the consumer would be able to, by taking advantage of their edge in processing speed and communication bandwidth. The different strategies are depicted in Figure 1 While an agent is “free” in terms of how to achieve its objective, the parameters described above suggest how it works. The crude heuristic used by the agents in the current version is: begin by offering the item at the desired price. If there are willing buyers, great. If not, as time goes along, lower the asking price to entice more interest. When the desired date to sell the good by rolls around, the asking price should be about the lowest acceptable price. Of course, all the interesting action is in the subtleties and nuances of how the selling agent goes about lowering the price. It is possible that there will be no buyers (perhaps the lowest acceptable price is too high, or no one interested in what the agent is selling). In this case, the agent fails to achieve its goal. The consumer can check on his/her selling agents, see which other agents they have talked to, and what prices they have been offered. This information might prompt the consumer to do something like lower an agent's price parameters, if they see that offers are coming in much lower than expected. The consumer always has final control over their agents. Then a selling agent reaches an agreement with a buying agent, their consumers may want to give the ok, so to speak, before the agents ``shake hands'' on the deal. To investigate the validity of our design we conducted an day-long experiment with people using the system to buy and sell real products. The next section describes this effort and discusses a few lessons learned.



.DVEDK$JHQW0DUNHWSODFH([SHULPHQW On October 30th 1996, the Media Laboratory organized a symposium on “Digital Life” for 171 attendees from industry [10]. Many of these people were not technically inclined and none had been introduced to the Kasbah concept beforehand. The attendees were given a total of three objects as well as some money (50 “bits”, as we called our unit of currency). We invited the attendees to participate in the experiment and to create “selling” agents to sell some of the objects they owned but did not want and to create “buying” agents to buy some of the objects they wanted to own. When a participant created an agent, he/she would determine its characteristics such as: is it a buying or a selling agent, what does the agent buy or sell (chosen from a limited list of goods), what is the initial asking (or offer) price, what is the final lowest asking (or final highest offer) price and what is the “strategy” the agent will use to lower (or raise) its price over time.

)LJXUH

3ULFHRI,WHPVRYHU7LPH

Attendees were able to create agents all day long. At the end of the day, they could exchange their money for wine (70 bits for one bottle). Hence, wine could be said to cor



respond to the “gold standard” of the economy created; people would decide how many bits objects were worth to them based on the fact that a bottle of wine was worth 70 bits. Participants in the agent marketplace experiment were immersed in an environment involving interactive kiosks, a transaction center, and several large-scale displays. participants created and interacted with their agents via 20 keyboardless kiosks which consisted of a computer workstation, monitor, mouse, and bar-code reader. Participants initiated a personalized session by scanning their name tag badge with the bar-code reader. Once automatically logged in, participants were provided with a list of their active agents, a list of completed transactions, and options for creating new agents and changing old ones. The interface was entirely mouse driven æ participants navigated a simple point-and-click interface to create, modify, and monitor their agents. No keyboard input was required. The interface was designed to facilitate short interactions, so that all participants could use the 20 kiosks during the 30 to 60 minute breaks. A typical participant session lasted about five minutes. In addition to the kiosks, several other devices and displays were used to disseminate information. Each participant received a personal alphanumeric pager which would notify them whenever one of their agents had made a deal. For example, if Andy had created an agent named James Bond to sell a lunch pail, and this agent then made a deal with one of Pattie's agents, Andy was paged with the message: “Andy, I sold your Media Lab lunch pail to Pattie Maes for $53. --James Bond”. Pattie was paged with a similar message from her agent. At one point, about halfway through the day, buying and selling agents sent suggestions to participants' pagers when it seemed unlikely that they would make a deal by the end of the day with the price parameters given by the participant. For example, if a participant created an agent to buy a box of chocolates and the maximum price the agent would offer was 20 percent lower than the average selling price of chocolates throughout the day, that agent would page the participant with the message: “Tip: you may want to raise my maximum offer if you want to buy those chocolates - James Bond.”



An 8 foot high projection display and two large monitors provided visualizations of marketplace statistics including histograms of buying and selling prices for each of the nine good types and a time-series projection illustrating the trend in actual transaction prices for each good over time. Figure 2 shows one of the time-series displays. A scrolling LED ’ticker-tape’ displayed up-to-the-minute market information about each of the goods being traded, such as the highest bid, lowest asking price, and last transaction amount. Upon being notified of a transaction (either via the pager or the participant interface), participants were told to report to the transaction center to exchange the good for the price agreed upon. We anticipated situations where one party would arrive several minutes after the other party (or fail to show up altogether) and wanted to avoid making the participants wait around since they had very busy schedules. Thus, when two participants arrived asynchronously, the transaction center, which had a supply of extra money and surplus goods, completed the one-sided deals. If the two participants arrived at the same time, they carried out the transaction privately between themselves. The transaction center was also used to provide general information and assistance, to make change, and to sell the wine at the end of the day. Two other incidents happened which can be identified in Figure 2. Around 12:30pm, one of the participants spread a rumor saying that at the end of the day, every person present would receive a free Media Lab watch. As a result the price of watches plummeted because people hurried to change the prices of their watch-buying agents. When it was pointed out later that this news was false, the price of watches picked up again. Another phenomenon that took place is that the prices of items generally rose towards the end of the day. This is the case because at the end of the day, people could only buy wine if they had more than $50. Anyone owning less than $50 was happy to spend all their money on whatever good they could buy, because the money became useless at the end of the experiment.



)LJXUH

1XPEHURI$JHQWVDQG'HDOVRYHU7LPH

Figure 3 shows the total number of agents, the number of active agents (those agents that have not yet made deals), and the cumulative number of completed deals made over the course of the day. The kiosks were accessed heavily during program intermissions and more moderately during lunch. This coincides with the number of agents spikes at the start of the experiment and at particular times - roughly between 11am and 12pm, 1pm and 2pm, and again at 3pm. The noticeable agents spike just before the 5:00pm market close was due to the transaction center infusing agents into the marketplace to distribute the remaining goods. By 1:30pm, the number of active agents in the marketplace began to decrease due to agents making deals faster than they were being created. This corresponds with the substantial rise in deals made around that 1:30pm time frame. The flat-line of agents and deals at 5:00pm is due to kiosks being switched off and the convergence of each agent’s negotiation price to its termination price at the 5:00pm hard-coded deadline. Finally, interviews with participants have revealed that the level of intelligence or sophistication which a buying/selling agent could ultimately demonstrate is limited more by user-agent “trust” issues than by limitations of AI technology. In particular, in order for these agents to be widely accepted, it is crucial that the agent's behavior can easily be understood and controlled by the user. 0RXNDV*XWWPDQ=DFKDULD0DHV,-(&6XEPLVVLRQ


3'$#6KRS So far electronic markets and physical markets have been totally separated. One of our other Kasbah-related projects, PDA@shop incorporates a layer of functionality that merges the virtual and the physical marketplaces for the benefit of the shopper. The consumer can use the Personal Digital Assistant (PDA) comparative shopping system to obtain alternative prices, as well as reputation ratings of the potential on-line seller. When the consumer finds himself in a physical vendor’s store he can use his PDA to communicate with Kasbah and create a buying agent. When the consumer creates a new broker-agent, he/she must set several parameters that the agent will use when it negotiates with the different selling agents. PDAs have several limitations that are relevant to resources required for comparison shopping, namely -- bandwidth, processing speed, and sustainable network connectivity. The goal of PDA@shop is to provide consumers with robust point-of-sale comparison shopping capabilities. Our prototype uses a commercial PDA with an intermittent network connection and the ability to spawn agents. The aim of the PDA@shop project is to build a domain-independent system, which uses minimum domain-specific knowledge. The system uses a Personal Digital Assistant (PDA) -- in our system we are using US Robotics Pilot Pro [12] -- as an interface communicating with a Kasbah Server for the back-end computational work in order to facilitate search for product types, their names and prices, as well as company names and related reputation quotes. The consumer can use his/her PDA while looking on a particular product in a shop, or a product advertisement, or even while browsing the web, in order to retrieve relevant information about the specific model and the producing company, as well as alternative brands available in the market and their respective prices and specifications. Thus the consumer increases his/her options, and can choose between the price offered in the physical market place or the alternative one on the electronic marketplace [13].


&XUUHQW(IIRUWV

&XUUHQW(IIRUWV Our current efforts try to address issues like distributed component-based marketplaces, merchant differentiation, value-based product comparisons, buying decision aids, visualization of marketplace data and activities, knowledge brokering, mobile shopping systems as well as trust and reputation of the second generation agent-mediated electronic commerce systems.

$&RQVXPHU%X\LQJ%HKDYLRU)UDPHZRUN Traditional Consumer Buying Behavior (CBB) models often need to be redefined to adapt to the electronic world [14], [15]. Figure 4 lists the six primary consumer buying behavior stages and shows where several representative agent systems fall within this space. Agent technology can apply to one or more stages of the CBB models, like Problem Recognition (using agent-based collaborative filtering tools, e.g. [16], [17]), Information Search and Evaluation of Alternatives (using constraint satisfaction technology, e.g. [18]), Negotiation and Purchase Decision (using negotiation distributed and integrative negotiation, e.g. [19], [20], [21] and [22]) and post-purchase evaluation (using reputation mechanisms, feedback, e.g. [24]).

)LJXUH

&RQVXPHU%X\LQJ%HKDYLRUZLWK$JHQW0HGLDWLRQ


&XUUHQW(IIRUWV

Briefly, the product brokering stage comprises the retrieval of information to help determine what to buy. This encompasses the evaluation of product alternatives based on consumer-provided criteria. The result of this stage is the ’consideration set’ of products. The merchant brokering stage combines this ’consideration set’ with merchantspecific information to help determine who to buy from. This includes the evaluation of merchant alternatives based on consumer-provided criteria (e.g., price, warranty, availability, delivery time, reputation, etc.). The negotiation stage is about how to determine the terms of the transaction. In traditional retail markets, price and other aspects of the transaction are often fixed leaving no room for negotiation. In other markets (e.g., stocks, automobile, fine art, local markets, etc.), the negotiation of price or other aspects of the deal are integral to product and merchant brokering

7HWHD7HWH7#7 0HUFKDQW'LIIHUHQWLDWLRQ T@T caters to the needs of both consumers and merchants by recognizing the importance of merchant differentiation and customer satisfaction in an ineffectively homogeneous and hostile retail environment. T@T’s approach engages consumer-owned shopping agents and merchant-owned sales agents in integrative negotiations to maximize their owners’ individual needs. Tete-a-Tete (T@T) differs from all other agent-mediated electronic commerce systems today. In addition to providing a more comprehensive shopping experience (e.g., by integrating three CBB stages), T@T also attempts to help merchants differentiate themselves along dimensions other than just price. For example, consumers are able to consider delivery time, return policies, support and repair services, loan options, brand and reputation, privacy policies, gift services, extended warranties, and other valueadded services in their buying decisions. T@T’s unique approach engages consumerowned shopping agents and merchant-owned sales agents in integrative negotiations to maximize their owners’ individual needs as shown in [23].


&XUUHQW(IIRUWV

)LJXUH

&RQVXPHURZQHG VKRSSLQJ DJHQWV LQWHJUDWLYHO\ QHJRWLDWH ZLWK PXOWLSOH PHUFKDQWRZQHG VDOHV DJHQWV

T@T uses an integrative negotiation protocol, which was inspired by argumentation theory as well as distributed constraint satisfaction research. Argumentation is comprised of proposals, counter-proposals, and critiques. At any moment, parties can either withdraw from negotiations or accept the current proposal. T@T provides an asymmetric version of argumentative negotiation as shown in Figure 6.

)LJXUH

&RQVXPHUVYLDWKHLUVKRSSLQJDJHQWV FULWLTXHSURSRVDOVIURPVDOHVDJHQWV

T@T sales agents automate the negotiation process for merchants. Shopping agents, on the other hand, actively assist shoppers during negotiations by providing them with a level of decision support to help them determine which merchant offering best meets their needs. This decision support is based upon multi-attribute utility theory (MAUT). 0RXNDV*XWWPDQ=DFKDULD0DHV,-(&6XEPLVVLRQ

&XUUHQW(IIRUWV

T@T’s MAUT mechanism provides shoppers with a real-time, utilitarian shopping experience. From a User Interface (UI) perspective, it is crucial that the consumer be able to easily express their preferences as well as being able to easily identify which product offering best meets his or her needs. If the information being displayed overwhelms the consumer, it could result in a dissatisfied customer and lost sales. Effective shopping interface design is a particular challenge because of the established norm for viewing search results - a textual list of names and numbers. In such an approach, shopping is cumbersome and confusing due to the need to scroll and click to see more results or modify preferences. Moreover, the relevance of each result item of the search to the consumer’s needs is not clearly conveyed (if at all). Today’s on-line cross-merchant shopping services do not provide adequate decision support with appropriate visual cues to indicate how much one option satisfies the consumer’s preferences nor how each option stacks up against one another. On-line retail shopping demands a user interface that helps consumers navigate an enormous sea of information to find the product offerings that best meet their individual needs. More specifically, an ideal shopping interface would help consumers decide which products to buy as well as which merchant to buy them from by considering each product’s full range of value. We foresee three main areas in user interface design necessary to provide this ideal shopping experience: capturing and uncovering consumers’ needs visualizing and ranking of product offerings mapping the integrative negotiations between consumers and merchants

.QRZOHGJH%URNHULQJDQG([SHUW)LQGHU These projects investigate how software agent technology can be used to facilitate knowledge brokering and “just-in-time-consulting”. Imagine a person trying to acquire expertise about a particular problem being able to get assistance in real time: we propose an infrastructure which allows that person to - without much effort - locate an


&XUUHQW(IIRUWV

expert who is available to help the expertise-seeker remotely at that moment in time. Specifically, we are building an “electronic market for expertise” by creating agents that can match up expertise-seekers and experts and make deals for an on-the-fly consulting relationship. To participate in this knowledge marketplace, a person can create selling agents that know what expertise that person has and when that person is available to help someone else (and possibly how much he charges). Someone in need for consulting can create a buying agent that knows what kind of expertise that person needs and by when (in the next 3 minutes, in the next 3 days, etc.). The buying agent may also have information on how to compare different people offering the expertise (criteria such as: price charged, reputation of the expert, level of expertise, availability, etc.). The market automatically matches up these different buying and selling agents. Once a match has been made, other media are used to implement the tutoring relationship (email, phone, as well as richer communication media). For instance, a doctor while examining a patient with a rare disease might want an additional expert's opinion for that disease in the next ten minutes (while the patient is still there). The doctor's agents would search for other doctors that can be of assistance in that case and based on a series of parameters (doctor reputation, availability, cost, expertise etc.) decide with which ones to begin negotiations with, conduct the negotiations along a set of different attributes (for instance, if the consultant doctor is in the middle of a meeting, the consulting fee would be higher), conclude them in real time and place a phone call to the consulting doctor. Our first approach to knowledge brokering is the Expert Finder [25]. The system’s first domain is that of knowledge about software applications: it is normal to find oneself “lost” in an application, not knowing how to get something done and unable to find it in the on-line help systems (when they exist), especially when one is new to the application. In these cases, the easiest way out is to ask someone who also uses this software, but has more experience with it (an expert). The present Expert Finder watches what programs are being used and makes that information available to the other Expert Finder agents. Each user's level of expertise is determined by a person's software usage 0RXNDV*XWWPDQ=DFKDULD0DHV,-(&6XEPLVVLRQ

&XUUHQW(IIRUWV

pattern. When help is needed using some software package, the user can query the agent, and it will find an expert to help with the chosen application. One of the most important brokering dimensions in this knowledge market is that of reputation. The next section describes our efforts to provide a robust, distributed mechanism for handling reputation.

&RQVXPHUWR&RQVXPHU5HSXWDWLRQ0HFKDQLVPV The members of electronic communities (like Kasbah [9], OnSale [27], or eBay [28]) are often unrelated to each other, they may have never met before and have no information on each other’s reputation. This kind of information is vital in Electronic Commerce interactions, where the potential counterpart’s reputation can be a significant factor in the negotiation strategy. In this project we propose two complementary reputation mechanisms that rely on collaborative rating and personalized evaluation of the various ratings assigned to each user. While these reputation mechanisms are developed in the context of electronic commerce, we believe that they may have applicability in other types of electronic communities such as chatrooms, newsgroups, mailing lists etc. Consumer to consumer electronic marketplaces introduce two major issues of trust among their users: The potential buyer has no physical access to the product of interest while he/she bids or negotiates. Therefore the seller could misrepresent the condition or the quality of his/her product in order to get more money.

The seller or buyer may decide not to abide by the agreement reached on the electronic marketplace asking at some later time to renegotiate the price, or even refusing to commit the transaction.

The reputation mechanisms used in both eBay and OnSale are insufficient because they ignore major problems introduced by the on-line nature of the electronic marketplaces. In on-line markets, it is relatively easy to adopt or change one’s identity. Thus, if a user ends up having a reputation value lower than the reputation of a beginner, he/she would have an incentive to discard his/her initial identity and start from the beginning. Hence, it is desirable that while a user’s reputation value may decrease after a transac-


&XUUHQW(IIRUWV

tion, it will never fall below a beginner’s value. We also want to make sure that even if a user starts receiving very low reputation ratings, he/she can improve his/her status later at almost the same rate as a beginner. If the reputation value is evaluated as the arithmetic average of the ratings received since the user joined the system, users who perform relatively bad at the beginning, have an incentive to adopt a new identity so that they get rid of their bad reputation history.Another problem is that the overhead of performing fake transactions in most electronic marketplaces is relatively low. Therefore two friends might decide to perform some dozens of fake transactions, rating each other with perfect scores so as to both increase their reputation value. It is desirable to have a limited memory of the users’ activity on the system long enough to increase the prediction power of the reputation values but short enough to reflect the latest behavior of the users. Even if we allow each user to rate another only once, another way to falsely increase one’s reputation would be to create fake identities and have each one of those rate the user’s real identity with perfect scores. A good reputation system would avoid both these problems. We have to ensure that those ratings given by users with an established high reputation in the system are weighted more than the ratings given by beginners or users with low reputations. We have built two reputation mechanisms, Sporas and Histos, that take into consideration all the above-mentioned problems and facilitate reputation brokering in agent mediated transactions. Sporas is a simple reputation mechanism which can be implemented irrespective of the number of rated interactions. Sporas borrows ideas from the Glicko chess rating system [29], extended to cover the weaknesses of our on-line systems. Sporas provides a reputation service based on the following principles: New users start with a minimum reputation value, and they build up reputation during their activity on the system. The reputation value of a user never falls below the reputation of a new user.


&XUUHQW(IIRUWV

After each transaction, the reputation values of the involved users are updated according to the feedback provided by the other parties to reflect their trustworthiness in the latest transaction. Two users may rate each other only once. If two users happen to interact more than once, the system keeps the most recently submitted rating. Users with very high reputation values experience much smaller rating changes after each update. This approach is similar to the method used in the Elo and the Glicko [29] systems for pairwise ratings. Each user has one reputation value, which is updated according to the formulae below:

1 Φ (5L )5LRWKHU (:L − ( (:L )) θ 1 Φ(5L ) = 1 − − ( 5 −1 − ' ) 1+ H σ 5 ( (:L ) = L −1 ' 5L = 5L −1 +

L

6SRUDVIRUPXODH

where θ is a constant integer greater than 1, Wi represents the rating given by the user i, Riother is the reputation value of the user giving the rating, D is the range of the reputation values, σ is the acceleration factor of the dumping function Φ (the smaller the value of σ, the steeper the dumping factor Φ), new users start with reputation equal to 0 and can advance up the maximum of 3000, the reputation ratings vary from 0.1 for terrible to 1 for perfect. Equation (1) shows that the change in the reputation value of a user receiving a rating of Wi from user Riother, is proportional to the reputation value Riother of the rater. In addition, since the reputation of a user in the community is the weighted average of


&XUUHQW(IIRUWV

non-negative values, it is guaranteed that no user can have a negative reputation value, thus no user can have a rating value lower than that of a beginner. Also, the weighed average schema guarantees that no user exceeds the maximum reputation value of 3000. However, if a user has a persistent real reputation value, iterations of Equation 1 over a large number of ratings will give an estimate very close to that value.

)LJXUH

&KDQJHRIUHSXWDWLRQIRUGLIIHUHQWXVHUVRYHUUDWLQJVZLWKθ

The expected rating of a user is expressed as the current reputation value over the maximum reputation value allowed in the system. Thus if the submitted rating for a user is less than his/her expected rating value, the reputation value of the user decreases.


&XUUHQW(IIRUWV

The value of θ determines how fast the reputation value of the user changes after each rating. The larger the value of θ, the longer the memory of the system. Thus, just like credit card history reports, even if a user enters the system with a very low reputation, if his/her reliability improves, his/her reputation value will not suffer forever from the initial poor behavior. Histos is a more complex reputation mechanism that assumes that the system has been somehow bootstrapped (for instance by using Sporas) so that there is an abundance of rated interactions to create a dense web of pairwise ratings. The system allows the users to evaluate the reliability of their counterparts in a personalized fashion. We can represent the pairwise ratings in the system as a directed graph, where nodes represent users and weighted edges represent the most recent reputation rating given by one user to another, with the arrow pointing towards the rated user. If there exists a connected path between two users, say from A to AL, then we can compute a more personalized reputation value for AL. When user A submits a query for the Histos reputation value of user AL we perform the following computation: The system uses a Breadth First Search algorithm to find all the directed paths connecting A to AL that are of length less than or equal to N. As described above we only care about the chronologically θ most recent ratings given to each user. Therefore, if we find more than θ connected paths taking us to user AL, we are interested only in the most recent θ paths with respect to the last edge of the path. We can evaluate the personalized reputation value of AL if we know all of the personalized reputation ratings of the users before AL in the path. Thus, we create a recursive step with at most θ paths with length at most N-1. If the length of the path is only 1, it means that the particular user, say C, was rated by A directly. The direct rating given to user C is used as the personalized reputation value for user A. Thus, the recursion terminates at the base case of length 1.


&XUUHQW(IIRUWV

For the purpose of calculating the personalized reputation values, we use a slightly modified version of the reputation function described above. For each user Ak, with m connected paths (of ratings higher than average) going from A to Ak, we calculate the reputation of Ak as follows: Let Wjk(n) denote the rating of user Aj for user Ak(n) at a distance n from user A1, and Rk(n) denote the personalized reputation of user Ak(n) from the perspective of user A1. At each level n away from user A1 the users Ak(n) have a reputation value given by:: 5

N

(Q ) =

( (Q − 1 ) • : (Q )) Ñ

'•Ñ 5

M

(Q ) P = deg ( $ (Q )) = : (Q ) θ ’ = min (θ , P ) 5 (1 ) = ' • : 1 (1 ) ∀ MN , ∃ :

MN

5

M

(Q − 1 )

MN

N

N

MN

N

+LVWRVIRUPXODH

where deg (Ak(n)) is the number of connected paths from A1 to Ak(n). The users Ak(n) who have been rated directly by user A1 with a rating W1k(1) have a reputation value equal to Rk(1).

Histos needs a highly connected graph. If there does not exist a path from A to AL with length less than or equal to N, we fall back to the simplified Sporas reputation mechanism. 6KRSSLQJRQWKH5XQ Shopping on the Run is extending our PDA@Shop and Kasbah projects in several ways. The goal of the project is to create an electronic market for goods and services, which would be accessed by mobile devices as well as from the desktop. Let us describe the functionality of the system by providing a few scenarios of consumers equipped with a PDA-like device with communication and positioning abilities:


&XUUHQW(IIRUWV

Walking near a bookstore: As a consumer is walking in Boston’s Newbury street, the shopping agent in her PDA is looking for deals from nearby merchants. Her profile information shows that she is interested in buying a Pablo Neruda book. The shopping agent starts negotiating with a couple of near-by bookstores for the book and closes a preliminary deal. At the same time, it checks the prices at different on-line merchants: the book is $2 cheaper (including shipping and handling), but the agent takes into consideration her preference for visiting bookstores, closes the deal and notifies her. Driving down a highway: A couple is driving down the highway and they are interested in a nice seafood restaurant in New Jersey that serves fresh catfish and can accommodate them in the next fifteen minutes. Again, the shopping agent contacts local restaurants and restaurant recommendation guides in order to best accommodate the request. An expert is nearby: A travel-lover is departing for the Aegean island of Myconos in a few of days and would like to talk with some sort of expert in the night-life and accomodations options on the island. He is willing to spend $20 to talk for a few minutes with an expert in Myconos if there was one in the vicinity in order to get some advice. The system provides just-in-time brokering for products, services, information and knowledge in a localized fashion. It can locate local merchants by using geography as a self-organizing merchant directory (without the need of a centralized yellow pages-like structure.) It can subsequently search and negotiate about products of interest to the consumer and receive information and locality-based services from content providers. In terms of equipment and infrastructure some sort of computational device (PDA-like or integrated to a mobile telephone,) with communication (multi-hop network, cellular service, two-way paging, etc.) and positioning capabilities (GPS, cellular telephone localization) is required. In our implementation we are borrowing elements from almost all the projects mentioned in this paper. We are currently testing the validity of the physical topology brokering, mobility and communication aspects of the architec-


&RQFOXGLQJ5HPDUNV

ture using a real-time simulator (described in detail in our CarTalk and Trafficopter [30] projects.)

&RQFOXGLQJ5HPDUNV Our first-generation agent-mediated electronic commerce systems are already creating new markets (e.g., low-cost consumer-to-consumer) and beginning to reduce transaction costs in a variety of business models. However, we still have a long way to go before software agents transform how businesses conduct business. This change will occur as Software Agent technology matures to better manage ambiguous content, personalized preferences, complex goals, changing environments, and disconnected parties, but more importantly, as standards are adopted to succinctly and universally define goods and services, consumer and merchant profiles, value added services, secure payment mechanisms and reputation.

5HIHUHQFHV [1]

Seattle Times, 1996. URL:

[2]

AdWorld, 1996. URL:

[3]

Infomaster, 1996. URL:

[4]

BargainFinder URL:

[5]

Jango URL:

[6]

J. Bosak. "XML, Java, and the Future of the Web." March 1997.

[7]

“Extensible Markup Language (XML).” World Wide Web Consortium (W3C) Working Draft, Nov. 1997.

[8]

Summer Institute of Linguistics XML URL:


5HIHUHQFHV

[9]

A. Chavez, and P. Maes. “Kasbah: An Agent Marketplace for Buying and Selling Goods” Proceedings of the First International Conference on the Practical Application of Intelligent Agents and Multi-Agent Technology (PAAM'96). London, UK, April 1996.

[10]

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.

[11]

P. Maes Agents that Reduce Work and Information Overload. Comm. of the ACM Vol. 37, No. 7, 1994

[12]

PalmPilot URL:

[13]

G. Zacharia, A. Moukas, R. Guttman, and P. Maes. "A PDA-based Agent System for Comparison Shopping." Proceedings of the European Conference on MM & E-Commerce, Bordeaux, France, September 1998.

[14]

I. Terpsidis, A. Moukas, B. Pergioudakis, G. Doukidis, and P. Maes. “The Potential of Electronic Commerce in Re-Engineering Consumer-Retail Relationships through Intelligent Agents.” J.-Y. Roger et al. (eds.), Advances in Information Technologies: The Business Challenge, IOS Press, 1997.

[15]

R. Guttman, A. Moukas and P. Maes. “Agent-mediated Electronic Commerce and Consumer Buying Behavior”, invited paper, Knowledge Engineering Review Journal, Vol 13:3, June 1998.

[16]

Firefly URL:

[17]

U. Shardanand and P. Maes. Social Information Filtering: Algorithms for Automating 'Word of Mouth'.” Proceedings of the Computer-Human Interaction Conference (CHI-95), Denver, May 1995.

[18]

PersonaLogic URL:

[19]

J. Rosenschein and G. Zlotkin. Rules of Encounter: Designing Conventions for Automated Negotiation among Computers. MIT Press, 1994.

[20]

R. Lewicki, D. Saunders, and J. Minton. Essentials of Negotiation. Irwin, 1997.

[21]

P. Gmytrasiewicz, E. Durfee, and D. Wehe. “A Decision-Theoretic Approach to Coordinating Multiagent Interactions” In Proceedings of the Eleventh International Joint Conference on Artificial Intelligence, August 1991.

[22]

R. Keeney and H. Raiffa. Decisions with Multiple Objectives: Preferences and Value Trade-offs. John Wiley and Sons, 1976.

[23]

R. Guttman and P. Maes. "Agent-mediated Integrative Negotiation for Retail Electronic Commerce." Proceedings of the Workshop on Agent Mediated Electronic Trading (AMET'98). May 1998.


5HIHUHQFHV

[24]

G. Zacharia, A. Moukas, and P. Maes. "Collaborative Reputation Mechanisms in Online Marketplaces" Proceedings of Hawaii International Conference on System Sciences, Wailea Maui, Hawaii, January 1999.

[25]

A. Vivaqua. The Expert Finder: Agents to assist in finding help. Working Paper, MIT Media Laboratory.

[26]

R. Doorenbos, O. Etzioni, and D. Weld. “A Scalable Comparison-Shopping Agent for the World Wide Web.” Proceedings of the First Int’l Conference on Autonomous Agents. Marina del Rey, CA, Feb. 1997.

[27]

OnSale Exchange. URL:

[28]

eBay. URL:

[29]

M. Glickman, "Parameter estimation in large dynamic paired comparison experiments”, Applied Statistics, 1998.

[30]

A. Moukas, K. Chandrinos, P. Maes. Trafficopter: A Distributed Collection System Traffic Information. Proceedings of the Cooperative Information Agents-98. In Lecture Notes in Artificial Intelligence. Kluwer Academic Publishers, 1998
