A Matchmaking Component for the Discovery of ... - CiteSeerX

A Matchmaking Component for the Discovery of Agreement and Negotiation Spaces in Electronic Markets Michael Ströbel and Markus Stolze IBM Research, Zurich Research Laboratory {mis,mrs }@zurich.ibm.com

Abstract This paper presents the design of an extended matchmaking component for electronic markets, which is able to identify negotiable agreements and the issues that are subject to the negotiation, in the case where basic matchmaking fails to find agreements that satisfy the constraints of the seller and the buyer. The foundation for this functionality is the introduction of negotiable constraints within the offer specification process. The extended matchmaking component complements our SILKROAD design and implementation framework for electronic negotiations. This framework also features other negotiation service components such as a mediation service, which may use the feedback from the extended matchmaking operation on agreement candidates and negotiation spaces, to suggest fair agreements on the basis of the Adjusted Winner algorithm for dispute resolution.

versa. The transaction, the object(s), and the involved agent(s) can be described with sets of properties such as the delivery date for the transaction, the colour of the object or the location of the agent. A property has a value domain with one or more values such as 12-12-00, green, or Switzerland. The seller advertises, for instance, to sell an object of the type notebook computer with the properties CPU = 600MHZ, RAM = 128MB, and disk = 6GB. Additional properties for the transaction may be price = $2000 and delivery time > 10 days. A buyer on the other hand might offer to buy a notebook with the property value domains RAM > 64MB, disk > 9GB, and price > $2000. Thus, both the seller and the buyer specify a set of constraints towards the value domains of the properties. In this example the offer of the seller matches that of the buyer and a transaction can be performed because compatible property values can be found that satisfy the constraints of both offers, e.g. RAM = 128 MB.

1. Introduction Let us assume a buyer wants to purchase an object using an electronic market, but cannot find an offer, which matches completely the requirements. Given this situation the buyer can, in general, either search for the object in other markets, or change the requirements and/or enter negotiations with sellers in order to make them change their offers. In this paper we demonstrate how negotiations in electronic markets can benefit from a support mechanism for the matchmaking phase preceding potential negotiations, which identifies agreement candidates and structures the subsequent negotiation by reducing it to a discussion of the negotiation issues discovered within the matchmaking operation. Hence the buyer can, in principle, avoid long haggling processes as well as negotiations with sellers that have a low agreement potential. An electronic market platform usually requires buyers and sellers to exchange offers-to-buy and offers-to-sell. The goal of this exchange is to reach an agreement on the performance of a transaction between buyers and sellers. A transaction transfers one or more objects (e.g. a product, money, etc.) from one agent to another agent and vice

Type Transaction Properties

Constraints

Offer-to-sell

Transaction Item Object Properties Properties Properties

Matchmaking Component

Agent Agent Properties Properties

Constraints

Offer-to-buy

Figure 1: Matchmaking overview. Hence, it is possible to formulate a symmetric constraint satisfaction problem with buyer and seller constraints (see Figure 1). This constraint satisfaction problem (CSP) can be solved automatically and the mechanism to find a solution (a set of compatible property values) for this CSP is denoted a matchmaking component. Typically, a matchmaking component will return a successful match if,

2

for each property specified in the offers, compatible property values can be found (see next section). In general, this matchmaking operation is performed for a set of offers with the goal to determine the set of successful matches. The outcome of ‘no successful match’ is particularly unsatisfactory in such cases, where values for perhaps only one property were incompatible, and that, by only a marginal property value difference. In the previous example, one additional constraint of the buyer specifying a minimum speed of 666MHZ for the CPU of the notebook would have resulted in incompatible values for the CPU property and accordingly, to ‘no successful match’ with the offer-to-sell. But maybe speed is not that important to the buyer, or it’s not critical that the speed is at minimum 666MHZ, or maybe the buyer would agree to a lower speed if the price were lower as well? The objective of this paper is to present an extended matchmaking component (EMC), which is able to identify negotiable agreements in the case where genuine constraint matching does not provide (enough) successful matches and where buyers and/or sellers are willing to relax some of their constraints. For this purpose, we explain in more detail in Section 2 the intuition of our solution comprising the symmetric preference specification, the matchmaking operation, the consecutive agreement candidate analysis, and the modes of operation. Section 3 outlines the application of the EMC in a larger context, namely in relation to other negotiation service components. Finally, in Section 4, we compare our solution with related work and discuss potential extensions.

2. The extended matchmaking component In this section, our novel matchmaking component is illustrated in detail. The contribution comprises the introduction of negotiable constraints, a basic matchmaking operation processing on the level of types and properties, and an extended matchmaking operation, which discovers different classes of agreement candidates and leverages further candidate analysis. To introduce the concepts of the extended matchmaking component, buyers and sellers are referred to as agents. The agent triggering the matchmaking operation is denoted the initiator. The initiator could also be a third party such as the market operator. Offers always have an associated type, which is characterised by a unique name and a set of common transaction, object, and agent properties. These properties can be used for the constraint specification. The originator of an offer always has one agent role, buyer or seller, and thereby declares the offer to either an O2B or O2S. An offer might specify constraints for the role of the originator (e.g. buyer) as well as for the role of the transaction counterpart, who assumes the opposite role (e.g. seller) and is the

originator of a complementary offer for the same type. Constraints for both agent roles are, for instance, common in the insurance industry where quotes in an O2S are dependent on the age, medical record, driving experience, etc. of the buyer.

2.1. Symmetric preference specification The first step requires agents to specify their preferences, which are represented as constraints defining acceptable or desired transaction, object, or agent property value domains. Unary constraints are expressed for one property (e.g. price > $2000). Binary constraints express relations between two properties on the basis of domain operators (,=,J,P,>). In the example, the seller might specify that payment take place before delivery of the notebook (e.g. payment_time < delivery_time). Unary and binary constraints can be combined with join operators (AND, OR). Hence, an agent could formulate a constraint for the notebook object, defining that the mouse device equals either a stick pointer or a touch pad. The constraints defining the value domain for one property are referred to as domain constraints. One property can have several domain constraints. Both buyers and sellers may use an additional structural element to characterise constraints – the negotiable tag. Constraints tagged to be negotiable are denoted ‘soft’ constraints, in contrast to ‘hard’ constraints that are not negotiable. With a negotiable constraint, an agent can express a continued interest in an agreement with the counterpart, in the case that no successful match can be found, and on the basis of a relaxation of this constraint. Of course this does not make sense for all properties (the agent location property for instance, will usually not be negotiable). The negotiable tag does not specify to what extent and whether at all a constraint might be relaxed should a conflict arise. It does not require the agent to actually change the negotiable constraint; it only indicates an agent’s willingness to do so. The actual decision to compromise might be subject to various considerations, such as the potential compensation regarding other properties etc. Upon completion of the preference specification, the offer specifications are submitted to the matchmaking operation. Consecutively, an offer can be subject to several matchmaking operations.

2.2. Matchmaking operation If the initiator triggers matchmaking, the EMC tries to find a solution to the constraint satisfaction problem, given the constraints of the initiating offer and the complementary offer. The basic matchmaking operation proceeds for a pair of one O2B and one O2S in the following way:

3

x For each property defined in the type for the transaction, object(s), and agent(s), the property value domains defined by the domain constraints in the O2S and O2B are determined. x Those property values that are part of both property value domains, constitute the compatible property value domain (CPVD). x If the CPVD is empty, the domain constraints in both offers are violated and the property represents a match conflict. A violation is always mutual. x If the CPVD contains at least one value, this property was matched successfully. x If all properties can be matched successfully, the original pair of O2B and O2S constitutes a successful match. To give an example for the violation of constraints, let us assume a buyer specifies an O2B with a constraint price < $2000. A corresponding O2S might define price = $2100. In this case the constraints in the O2B and O2S are violated because $2100 does not intersect with the value domain < $2000. If the seller defined price > $1800, the property would match successfully, but result in an agreement value range of $1800 < price < $2000. This means that for this pair of offers, the buyer and the seller still have to agree on a specific value within the agreement value range. An agreement value range does not have to be steady but can also feature a discrete set of values (e.g. {black, white}). Binary constraints are violated if the defined domain operation (e.g. ‘smaller than’) fails for all combinations of values belonging to the respective property value domains of the complementary offer. This can also be illustrated in the sample constraint payment_time < delivery_time in an O2B. If the O2S specifies property value domains of t + 2 < payment_time < t + 5 and delivery_time = t + 3 both properties would match successfully. What happens if no constraints, or constraints from only one agent are defined for a property? To resolve such nullproperties the following approaches can be suggested: x This situation is generally avoided. The matchmaking operation requires the agents to specify constraints for all properties in the type. x The matchmaking operation requires at least one domain constraint from either agent for all type properties. If only one agent expresses a domain constraint, two options exist: the property is matched successfully and the CPVD is defined by this domain constraint, or the property creates a match conflict. x The matchmaking operation is restricted to the properties with domain constraints and the nullproperties are marked as open issues. x The type definition may include default property values, representing best practices or common trade

standards. These default values are initially suggested for the null-properties. Using the foundation of the basic matchmaking operation, we propose an extended matchmaking operation, which takes the notion of negotiable constraints into account: 1. If the O2B and O2S constitute a successful match, the EMC determines whether one or more successful property matches result in agreement value ranges. If not, the O2S is added to the matching agreements set (MAS). If yes, the O2S is moved into the distribution agreements set (DAS). 2. If one or more properties created matchmaking conflicts, the EMC checks for each property, whether the violated domain constraints are negotiable. 3. If both the seller’s and buyer’s constraints are negotiable, the constraints are marked as double-sided negotiable constraint (DSNC) and the O2S is moved to the negotiable agreements set (NAS). 4. If this is true for only one agent, the negotiable domain constraints of this agent are marked as singlesided negotiable constraint (SSNC) and the O2S is also moved to the NAS. 5. In the case where neither the buyer nor the seller issued negotiable constraints for the property creating a match conflict, the extended matchmaking operation fails for this pair of O2B and O2S. In principle, the basic and extended matchmaking operation can be executed on a set of n initiating offers and m complementary offers, thus forming n x m match candidates (see Section 4.2). The result of the EMC operation is then a classification of the match candidates into four sets (see Figure 2):

Constraints

Basic matchmaking operation

Offer-to-sell

Constraints

Offer-to-buy

Extended matchmaking operation

Matching agreements MAS

Distribution agreements DAS

Negotiable agreements NAS

Unsuccessful matches

Figure 2: Extended matchmaking result sets

4

x A set of match candidates where no constraint violations occurred and the CSP solution contains only compatible property value domains with single values (MAS). x Match candidates where no constraint violations occurred but the solution contains agreement value ranges (DAS). x Match candidates where negotiable constraints from the buyer and/or the seller were violated (NAS). x A set with unsuccessful matches. All match candidates in the MAS, DAS, and NAS are referred to as agreement candidates. Depending on the configuration of the matchmaking operation, agreement candidates with open issues (see above) might be identified and classified as well. The discovery of open issues is also a valuable input for the analysis of subsequent agreement processes. However, a complete discussion of this additional aspect would exceed the scope of this paper.

2.3. Agreement candidate analysis Each match candidate, i.e. each O2B and O2S pair, in the MAS, DAS or NAS defines an agreement candidate. The goal of the agreement candidate analysis is to derive conclusions for the potential subsequent agreement process from the agreement candidate set classification and the constraint marking. This agreement process is characterised and constrained by an agreement scenario, which defines the roles and protocols as well as the context of the process. 2.3.1. Matching or distribution agreements For agreement candidates in the MAS a subsequent negotiation is not necessary. These offers did not violate any of the constraints of the initiator and accordingly, an agreement is already reached. To perform the associated transaction, the remaining agreement scenario is reduced to the choice of one successful match in cases where the MAS contains more than one candidate. For the DAS a conflict of distribution can be defined. In principle, no constraints are violated, but to reach a final agreement, the agents have to agree on single property values for all agreement value ranges identified in the CSP solution. If the EMC returned a range of $1800 < price < $2000, the agents have to agree on a price within that range. The total of all identified agreement value ranges for an O2B and O2S pair defines the agreement space. The resulting agreement scenario is characterised by the desire of each agent to achieve optimal property values within that space, on the basis of their preference structure. This is not a negotiation problem in the narrow sense, because, in principle, all constraints of the agents were satisfied and therefore any solution should be acceptable. However, the agents still face a conflict of surplus distribution, comparable to a situation of overlapping reservation

prices. In practice, agents might even risk to miss a viable agreement in order to maximise their returns. One approach to this configuration problem is to use joint maximum utility gain as the optimisation criterion (see for example [18]). 2.3.2. Negotiable agreements True negotiation processes, requiring compromises from at least one agent in order to reach an agreement are necessary for the offer pairs in the NAS. The required negotiation process can be further analysed on the basis of the marking performed during the extended matchmaking operation. Those properties that created a match conflict for a pair of offers in the NAS are referred to as negotiation issues. The semantics of the associated domain constraint marks are as follows: x Single-sided negotiable constraint Only the originator of this constraint is flexible towards this issue. The corresponding domain constraint(s) in the complementary offer is (are) not negotiable. x Double-sided negotiable constraint Both agents declared flexibility towards this issue. Constraints originally tagged to be negotiable, but not marked within the extended matchmaking operation are not relevant in the agreement candidate analysis. The flexibility indicated by the agents was not operative (used) in the matchmaking operation as no actual violation occurred for these constraints. Each offer in an agreement candidate pair in the NAS may comprise single-sided as well as double-sided negotiable constraints. This result is denoted a negotiation signature. The combination of the two negotiation signatures within an agreement candidate defines the negotiation space. A negotiation space comprises the issues, the value boundaries for these issues, and the constraint markings. The complexity of the agreement process for a negotiation space can also be assessed on the basis of the following primary constellations: x Single-sided negotiation space Only one offer features negotiable constraints, in this case single-sided negotiable constraints. x Double-sided negotiation space Both offers contain negotiable constraints. For single-sided negotiation spaces an agreement depends on the flexibility of one agent. The constraints, which caused a violation, are tagged to be negotiable and are marked to be operative. Thus an agreement can be reached if the agent relaxes these constraints accordingly. For instance, if a seller offered guarantee = 6 months, an agreement is possible with the buyer if the resulting property value domain is changed to guarantee = 12 months. Double-sided negotiation spaces, on the other hand, require mutual concessions in order to reach an agreement.

5

One possibility to further extend the analysis of the negotiation space is to assess the number of negotiable issues for each agreement candidate. One can distinguish singleissue and multi-issue negotiation spaces. A double-sided negotiation space can still be of single-issue nature – if both offers contain only one double-sided negotiable constraint. Another element of the analysis is to determine the range of disagreement for an issue by evaluating the difference between the boundaries of the property value domains for the negotiation issues. If the O2S specifies price = $2100 and the O2B’s constraint is price < $2000, then the range of disagreement is $100. This is obviously less meaningful for non-metric issues such as colour. The extent to which agents are willing to negotiate certainly depends on this property value boundary difference. But further analysis of the negotiation space on a general level is not possible because the importance of the issues to the buyer/seller and the range of intended or acceptable flexibility will vary from case to case. If the seller only needs to relax one constraint, but this issue is crucial to the seller and small concessions may already manifest a big revenue loss, then the probability of an agreement might be lower than in a case with multiple issues of minor importance and greater boundary differences. Finally, the subsequent agreement scenario can also be characterised on the basis of a set-level analysis. Let us assume the NAS comprises multiple agreement candidates with single-sided negotiation spaces where only the buyers need to be flexible. This NAS structure suits well the requirements of an auction agreement scenario. Depending on the number and homogeneity of the issues a single- or multi-issue auction service [2] can support the subsequent agreement process. In a mirrored single-sided NAS structure a set of sellers may have offers with only the price property operative negotiable, thus a reverse single-issue auction can be initiated. Comparing the different result sets might also provide additional feedback. There could be, for instance, only one agreement candidate in the MAS, but several candidates in the NAS. Whereas an agreement for the candidate in the MAS is straight forward, the effort invested in negotiating additional agreements with the candidates in the NAS might be well-invested, if these agreements exploit winwin tradeoffs, and thereby achieve overall higher benefits for the initiator than the agreement in the MAS.

2.4. Modes of operation In general the EMC will return the solution(s) to the CSP problem in terms of the four offer sets introduced. Regarding the structure of the NAS, the matchmaking operation and the type of the output can be further controlled by, for instance, the matchmaking initiator or the

market operator. This control is performed on the basis of policy properties. Potential policies for the basic matchmaking operation regarding unspecified property value domains in offers were already addressed in Section 2.2. One policy option for the extended matchmaking operation that allows controlling the complexity of the negotiation space, is that only agreement candidates with a maximum of n negotiation issues may belong to the NAS. For n = 1, only agreement candidates with single-issue negotiation spaces will classify for the NAS. A second policy option is that the EMC provides feedback on constraint violations for one offer regarding the set of complementary offers. The EMC can determine those constraints that caused the most violations. It can also identify those constraints, where a change to a soft constraint specification would result in x additional agreement candidates. Finally, the third option, which only applies to doublesided negotiation spaces, is that the EMC uses importance ratings of the sellers and buyers for negotiable constraints in the matchmaking process. If the initiator were only interested in agreements with win-win potential, which require tradeoffs (see for example [11]), the EMC would only classify agreement candidates for the NAS with more than one negotiation issue and different preference distributions of the agents towards these issues. A conflict where for the buyer delivery time is much more important than price, where as the seller favours the price over the delivery time, would constitute such an example.

3. Integrated agreement process support This section shifts the focus from details regarding the operation of the extended matchmaking component to the larger context of matchmaking in electronic markets, namely the integration of the EMC with other services supporting the agreement process of buyers and sellers. After the analysis of the agreement candidates, one possibility for the agents involved is to engage in conventional bilateral negotiations on the basis of the disclosed agreement and negotiation spaces. The alternative is a continued intermediation of the agreement process through additional services. The configuration of these services can be based on the agreement candidate analysis. To support a broad range of agreement scenarios, the service operations and the interaction of these services have to be configurable. Another requirement is a common language for the interaction of various support services– not only for the aspects of what is subject to the agreement, but also regarding how the agreement process is proceeded (because this depends on the identified agreement scenario). Our SILKROAD framework (see [16] and [17] for details regarding SILKROAD) provides this flexible support platform and common negotiation language. Hence, to outline

6

the options for the consecutive agreement process support, first the overall architecture of the SILKROAD framework is outlined, then a more detailed example for the interaction between the EMC and another negotiation service component is given.

3.1. Framework architecture The primary goal for the SILKROAD framework is to facilitate the design and implementation of electronic negotiation support. The two core elements of SILKROAD are the ROADMAP and the SKELETON. The SKELETON provides several modular and configurable negotiation service components and can be classified as an application framework [4]. One particular electronic negotiation support system is a customised instance of the SKELETON framework. It may support one or multiple agreement scenarios. The customisation affects the runtime behaviour of the instance and is based on specifications expressed in the SILKROAD design phase. This design phase is structured according to the ROADMAP action model. Core components of the SKELETON (see Figure 3) are the agent manager (AM), the policy manager (PM), and the negotiation service components (NSC), which assist or automate a task in the agreement process. The agent manager handles all interactions with agents, which can, for instance, submit an offer or initiate the matchmaking process.

Agent Agent Agent

Communication Repository

Agent Manager

Match

Bundle

Offer Repository

Policy Manager

Bid

Organisation Repository

Score

Mediate Contract

Figure 3: Framework architecture overview. The actual invocation of negotiation service components is performed by the policy manager on the basis of the negotiation design, i.e. it’s runtime representation, which is persisted in the organisation repository. This repository holds a state machine which defines, depending on the current state of the negotiation and the event raised by an agent, the next negotiation service component to be invoked. The choices of policy options introduced for the

operation of the EMC (see Section 2.4) are also part of this state machine specification. The communication repository defines the syntax and semantics for the offer specifications, i.e. the transaction, object, and agent properties. All NSCs operate on sets of offers, which are represented as XML documents in the offer repository. Upon invocation, the EMC, like any other NSC, retrieves the offer XML documents stored in the repository database, parses the documents and then performs the matchmaking operation on the basis of the retrieved offer data (constraints and properties). The result sets of the matchmaking operation (MAS, DAS, NAS) are stored together with a stateful representation of the service, which can be reactivated any time. If offers were modified within the operation (e.g. through the marking process), the offers in the repository database are also updated.

3.2. Negotiation service component interaction Apart from the EMC presented in this paper and the Mediate NSC outlined in the next section, the following service components are planned to be available within the framework (see Figure 3, the EMC is indicated in this figure as the Match NSC, the Mediate NSC will be explained in the next section): x Score The score service receives a set of agreement candidate offers and an initiating offer specifying evaluation criteria (combinative and utility functions) to calculate a ranking for the set of candidate offers, depending on their evaluation scores, thus determining the ‘best’ offers from the perspective of the initiating agent (see [14]). x Bid Comparable to common electronic auction systems (see for example [9]), this service persists and advertises offers in a bidding round, if they comply with certain bidding rules (e.g. ‘ascending prices’), and until a clearing rule applies (expiration of the bidding period, time of inactivity etc.), and the winning bid is determined. x Contract Any offer can be transformed into a legally binding offer with the signature of the issuing agent. Depending on the agreement scenario chosen, a final contract might require that two matching offers, both signed by the originator with respect to the complementary offer be found (one-sided contracting), or that one offer is signed by both agents (double-sided contracting). This service component requests agents to sign offers within a certain timeframe, checks the validity of the signatures and the necessary authorisation, archives the contract, and initiates the enactment process.

7

x Bundle Based on this service, offers can be collected or aggregated to cumulative offers. A typical application is demand aggregation, where, for instance, buyer agents group together in order to increase their negotiation power and consequently, to achieve better prices [19]. Another type of bundling takes place if one offer-to-sell consists of a bundle of complementary services (e.g. transportation from A to B, B to C, and C to D) provided by different seller agents. Though the usage of each of these four services for the support of electronic negotiations has already been evaluated in related research, the contribution of SILKROAD is the provision of these services together with the Match and Mediate NSC in one framework as well as the integration of these service implementations with a negotiation design approach. The design of a negotiation might specify an arbitrary sequence of NSC operations. To support, for instance, a multi-issue auction scenario, a typical sequence of NSC operations would be Match – Bid – Score – Bid – Contract. In this scenario, the matchmaking operation determines whether an offer suits the constraints of the initiating offer to participate in the auction, for instance, by being part of the MAS. If this is the case, the Bid NSC receives the offer and the Score service is used to determine whether this offer is better than the current best offer. If yes, the offer becomes the current best offer. After the clearing of the auction, this winning offer is then transformed into a legal contract by the Contract NSC. In the following section, a more detailed example for service interaction within the SILKROAD framework is outlined.

3.3. Interface example: Match/Mediate The Mediate NSC in the SILKROAD framework explicitly requires a precedent operation of the EMC (see [15] for a detailed discussion). The goal of the Mediate NSC is to suggest an agreement in cases where a conflict exists between a buyer and seller regarding the properties of the intended transaction or the associated transaction object(s). In order to suggest agreements with a high potential of acceptance, the Mediate NSC uses procedures for dispute resolution stemming from game theory, i.e. Proportional Allocation or Adjusted Winner [3]. An important feature of these procedures is, that they are not vulnerable to strategic misrepresentation, unless one agent has advanced information about the exact preferences of the other agent. Before dispute resolution can be applied, it is critical to determine the set of issues which define the conflict, and where both sides can accept that the other side wins  meaning that the final property value of this transaction,

object, or agent property, equals the suggested property value in the offer of the winning agent. This definition of issues is the output of the EMC operation: the negotiation spaces defined for the agreement candidates in the NAS. Offer-to-buy N: Price N: Payment N: Return Policy N: Delivery Time CPU RAM N: Disk Mouse

Offer-to-sell Double-sided negotiable constraints

Single-sided negotiable constraints

N: Price N: Payment N: Return Policy N: Delivery Time CPU N: RAM Disk N: Mouse

Figure 4: Constraint marking example. For double-sided, multi-issue negotiation spaces with double-sided negotiable constraints (where both agents declared flexibility towards an issue, see Section 2.3.2) the Mediate NSC is able to suggest an agreement, which by using the Adjusted Winner procedure, is efficient, equitable, and envy-free. For negotiation spaces with singlesided negotiable constraints (see Figure 4, ‘N:’ stands for ‘Negotiable’) this is not possible, because the outcome of the procedure for each issue might request either side to relax their constraint, but SSNCs indicate flexibility on only one agent’s side. For the DSNCs in Figure 4, the negotiation space could be defined by the EMC as follows: Issue Price Payment Return Policy Delivery Time Points gained

Buyer Constraint Weight < 2000 60 4 weeks 30 Full 5 1 week 5 90

Seller Constraint Weight 2400 10 On receipt 25 50% 30 > 2 weeks 35 65

Table 1: Negotiation space example. Using this input, the Mediate NSC requests the agents to assign importance ratings ranging from 0 to 100 to the four issues – the total of all ratings is 100. Given the values in the example, the Adjusted Winner procedure determines the suggested agreement property values initially to be price = 2000, payment = 4 weeks, return policy = 50%, and delivery time = 2 weeks. This initial distribution is not equitable and envy-free as the buyer gained more points than the seller. Hence, an additional equitability adjustment is suggested on the issue with the smallest preference ratio according to the following calculation: Payment mode: 65 + 25x = 90  30x Ö x = 5/11 Final split: 17 days Points achieved: |76.36 points

8

It is up to the buyer and the seller whether to accept this suggested agreement. The benefit of this solution is, that the procedure is fair (both agents have an interest in being honest) and long haggling processes are avoided. If the buyer and the seller accept this suggested agreement, the Contract NSC can be invoked.

4. Discussion This paper presented the design of an extended matchmaking component, which complements genuine constraint satisfaction with constraint relaxation analysis to not only discover matching pairs of offers within an electronic market, but also potential agreements, which require additional negotiations. The extended matchmaking operation classifies agreement candidates in three sets: matching agreements, distribution agreements, and negotiable agreements, where the area of conflict between a buyer and a seller is defined by the negotiation space. This negotiation space comprises the negotiation issues, the boundaries of these issues, and the areas of compromise. By means of an analysis of this conflict information, the complexity of a consecutive negotiation for the agreement candidate can be assessed. Hence, the potential negotiation is pre-structured and may be reduced to the essential discussion of the already identified negotiation issues. In addition, this paper demonstrated that the EMC provides a structured foundation for the application and customisation of complementary support services such as conflict resolution or contracting. The preference specification suggested for the EMC enables agents to control their level of compromise in a transparent and dynamic way. Depending, for instance, on the current market situation and the feedback from previous matchmaking operations, negotiable constraints can be activated and de-activated, thus resulting in more or fewer agreement candidates discovered by the EMC. Furthermore, a detailed agreement and negotiation space analysis provides the agents with additional information, which can be used on a case-by-case basis to assess whether they are willing to relax constraints regarding the operative set of negotiation issues for a specific agreement candidate. From the perspective of the operators of the EMC, various policy options for the customisation of the matchmaking operation can be used to regulate the market, for instance by controlling the number of potential agreement candidates.

4.1. Related work The EMC is based on the matchmaking concepts developed for the ViMP project [5]. Iwaihara [7] and Raman et al. [12] suggested similar basic matchmaking platforms, but do not elaborate on the usage of matchmaking for the preparation and support of agreement processes, which is

one of the main goals of this paper. On the other hand, negotiation support systems have to date focused on core bilateral or multilateral agreement processes such as conflict resolution (for an overview see [8] or [1]), but have not addressed the phase typically preceding a negotiation in electronic markets, namely the selection of the most promising agreement candidates. In addition to the application of matchmaking to electronic negotiations, novel to the EMC approach is, that constraint satisfaction based on symmetric preferences is augmented with symmetric indications of offer flexibility in order to identify spaces of agreement and disagreement, thus supporting the potential subsequent negotiation process between buyers and sellers. Typical autonomous agent systems, e.g. [20], rely on CSP approaches to resolve conflicts but require complete utility function specifications to resolve conflicts and have no means to express potential compromises. On the other hand, negotiation support systems such as Sardine [10] use notions of flexibility, but not in a symmetric way – only the buyer can indicate flexibility. Finally, a similar approach with flexible and symmetric constraints was suggested by Reeves et al. [13], but the proposed solution is focused on bilateral conflict resolution, not on the identification of the potential agreement candidates from a multilateral perspective. However, this approach could be represented as an additional SILKROAD negotiation service component that offers an automated conflict resolution method alternative to Adjusted Winner or Proportional Allocation.

4.2. Extension The current implementation assumes that the matchmaking process is initiated by one agent for one offer. Interesting opportunities arise if the matchmaking process is initiated centrally, comparable to the clearing process in an exchange. Then sets of agreement candidates with alternative initiating offers, the totality of matches, could be considered, thus allowing the introduction of market optimisation measures such as the overall welfare, and imposing competitive pressure on the initiating offers. The measure of the overall success of the matchmaking operation could be based, for instance, on the level of conflict, defined as either the total number of matchmaking conflicts or the total of buyer/seller utility disparities. Next to an evaluation of the extended matchmaking component in real world scenarios, we are currently investigating an extension for the EMC where a constraint can not only be tagged to be negotiable, but also be complemented by a certain value domain probability (distribution). In supply chain scenarios, it could be the case that a delivery time of 10 days, for instance, can only be guaranteed with a probability of 90% and in 10% of the cases, delivery is delayed. Matching with respect to these prob-

9

abilities can be incorporated into the EMC, resulting, for instance, in distribution agreement candidates with an associated cumulative level of uncertainty. We are also envisioning another extension of the matchmaking component, which not only matches offers of agents with the role buyer or seller, but also offers of third parties. The offers of third parties such as intermediaries, could define constraints for the properties of the other agent roles (e.g. ‘match only buyers older than 18 and sellers with a location in Switzerland’), as well as for other offer type properties. To keep the symmetry of the approach, the properties of this third party agent role might also be subject to constraints specified by the other agents (e.g. intermediary.fee < $100). The resulting three-waymatchmaking-relationship can be exploited to support specific requirements (e.g. access restrictions) of electronic market operators or other intermediaries.

Acknowledgements The authors would like to thank their colleagues Yigal Hoffner and Chris Kenyon at IBM’s Zurich Research Laboratory who stimulated some of the ideas for the extension of the extended matchmaking component, as well as the anonymous reviewers for viable feedback on the ideas presented.

References [1] Beam Carry, Segev Arie, Bichler Martin, Krishnan Ramayya ‘On Negotiations and Deal Making in Electronic Markets’, Information Systems Frontier Vol.1 No.3, 1999, pp.241-258. [2] Bichler Martin ‘Decision Analysis - A Critical Enabler for Multi-attribute Auctions’, Proceedings Twelfth International Bled Electronic Commerce Conference, Bled, Slovenia, 1999. [3] Brams Steven, Taylor Alan Fair Division: from cake-cutting to dispute resolution, Cambridge University Press 1996. [4] Fayad Mohamed, Schmidt Douglas, Johnson Ralph Building Application Frameworks - Object Oriented Foundations of Framework Design, Wiley, New York, 1999. [5] Field Simon, Hoffner Yigal ‘ViMP - A Virtual Marketplace for Insurance Products’, EM Journal, Vol. 8, No. 4, 1998. [6] http://www.javasoft.com/products/ejb/ - visited 03/12/01. [7] Iwaihara Mizuho ‘Supporting Dynamic Constraints for Commerce Negotiations’, Proceedings Second International Workshop on Advanced Issues of E Commerce and Web Based Information Systems. WECWIS 2000 pp.12-20. [8] Jelassi Tawfik, Foroughi Abbas ‘Negotiation Support Systems: An Overview of Design Issues And Existing Software’, Decision Support Systems Vol.5, 1989, pp.167-181. [9] Kumar Manoj, Feldman Stuart ‘Business Negotiations on the Internet’, Proceedings inet’98, Geneva Switzerland, 1998. [10] Morris Jane, Maes Patty ‘Negotiating Beyond the Bid Price’ Workshop Proceedings of the Conference on Human Factors in Computing Systems (CHI 2000), The Hague, Netherlands, 2000. [11] Raiffa Howard The art and science of negotiation, Harvard University Press 1982. [12] Raman Rajesh, Livny Miron, Solomon Marvin ‘Generic Matchmaking with Classified Advertisements’, First IAC Workshop on Internet-Based Negotiation Technologies, March 18-19, Yorktown Heights 1999. [13] Reeves Daniel, Grosof Benjamin, Wellman Michael, Chan Hoi ‘Automated Negotiations from Declarative Contract Descriptions’, Proceedings AAAI-2000 Workshop on Knowledge-Based Electronic Markets, Austin, TX, 2000. [14] Stolze Markus ‘Soft Navigation in Product Catalogs’ European Conference on Digital Libraries, Greece, 1998. [15] Ströbel Michael ‘A Framework for Electronic Negotiations Based on Adjusted Winner Mediation’, Proceedings of the DEXA 2000 Workshop on e-Negotiations, London, UK, 2000. [16] Ströbel Michael ‘Communication Design for Electronic Negotiations’, Proceedings of the WWW10 Conference, Hong Kong, 2001, pp.9-20. [17] Ströbel Michael ‘Design of Roles and Protocols for Electronic Negotiations’, to appear: Electronic Commerce Research Journal, Special Issue on Market Design, 2001. [18] Teich Jeffrey, Wallenius Hanelle, Wallenius Jyriki ‘Multiple Issue Auction and Market Algorithms for the World Wide Web’, Interim Report IR-98-109/December, International Institute for Applied Systems Analysis Laxenburg Austria 1998. [19] Tsvetovat Maksim, Sycara Katia ‘Customer Coalitions in the Electronic Marketplace’, Proceedings of the International Conference on Autonomous Agents, New York, 2000. [20] Willmott Steven, Calisti Monique, Faltings Boi, Santiago Macho-Gonzalez, Belakhdar Omar, Torrens Marc ‘CCL: Expressions of Choice in Agent Communication’, The Fourth International Conference on MultiAgent Systems (ICMAS-2000), Boston, MA, 2000.