Retailer selection in future open competitive communications environments

Computer Communications 25 (2002) 662±675

www.elsevier.com/locate/comcom

Retailer selection in future open competitive communications environments q M. Louta, P. Demestichas*, E. Tzifa, E. Loutas, M. Anagnostou Computer Science Division, National Technical University of Athens, 9 Heroon Polytechneiou Street, Zographou 15773, Athens, Greece Received 5 September 2000; revised 29 May 2001; accepted 18 June 2001

Abstract The highly competitive communications markets of the future should encompass mechanisms for enabling users to ®nd and associate with the most appropriate retailers, i.e. those offering at a certain time period adequate quality services in a cost ef®cient manner. This paper presents such mechanisms. Our starting point is the de®nition of a business case, through which the role of the best candidate-retailer selection problem is explained. In the sequel, the problem is analysed and the identi®ed sub-problems are concisely de®ned, mathematically formulated and solved. The identi®ed components of the best candidate-retailer selection problem involve the evaluation of the quality of a retailer offer and the reduction of the set of candidate retailers by exploiting learning from experience notions. In the ®nal sections, results are provided and concluding remarks are made. q 2002 Elsevier Science B.V. All rights reserved. Keywords: Retailer; Service architecture; Telecommunications information networking architecture; 0±1 Linear programming; Learning from experience

1. Introduction The ongoing liberalisation and deregulation of the telecommunication market will introduce new actors [1±3]. In principle, the main role of all players in such a competitive environment will be to constantly monitor the user demand, and in response to create, promote and provide the desired services and service features. The following are some key factors for success: First, the ef®ciency with which services will be developed. Second, the quality level, in relation with the corresponding cost, of new services. Third, the ef®ciency with which the services will be operated (controlled, maintained, administered, etc.). The challenges outlined above have brought to the foreground several new important research areas. Some of them are the de®nition of new business models, the elaboration on e-business concepts [4±7], the speci®cation of service architectures (SAs) [8±14], the development of advanced service creation environments (SCEs) [15±17] and service features (e.g. the personal mobility concept [9,18]), and the exploitation of advanced software technologies, (e.g. distributed object computing [19,20] and intelligent mobile agents q In part, this work has been supported by the Commission of the European Communities, within the ACTS Project MONTAGE (Mobile and Intelligent Agents in Accounting, Charging and Personal Mobility Support Ð AC325). * Corresponding author. Tel.: 130-1-772-25-29; fax: 130-1-772-25-34. E-mail address: [email protected] (P. Demestichas).

[21±24]). The aim of this paper is, in accordance with the cost-effective QoS provision and the ef®cient service operation objectives, to propose enhancements to the sophistication of the functionality that can be offered by service architectures in open competitive communications environments. A typical view of the competitive telecommunications world of the future can be the one depicted in Fig. 1. Without being exhaustive, ®ve main different entities can be identi®ed, namely, user, retailer, (third party) service (or content) provider, broker and connectivity provider. The role of the (third party) service (content) provider is to develop and offer services (content). The role of the retailer is to provide the means through which the users will be enabled to access the services (content) of (third party) service (content) providers. Limited by techno-economic or administrative reasons each retailer offers services only inside a domain. Moreover, it can be envisaged that an arbitrary area will, in general, fall into the domain of several retailers (Fig. 2). The broker assists business level entities in ®nding other business entities. Finally, the role of a connectivity provider is to offer the network connections necessary for supporting the services. Highly competitive and open environments should encompass mechanisms that will enable users to obtain services through the most appropriate retailers, i.e. those offering, at a given period of time, adequate quality services in a cost ef®cient manner. In this paper, the pertinent

0140-3664/02/$ - see front matter q 2002 Elsevier Science B.V. All rights reserved. PII: S 0140-366 4(01)00393-0

M. Louta et al. / Computer Communications 25 (2002) 662±675

663

Fig. 1. A view of the business level entities in the future competitive telecommunications environment.

problem is called best candidate retailer selection. The aim of this paper is (primarily) to address this problem from one of the possible theoretical perspectives and (secondarily) to show how the solution can be incorporated in service architectures that run in the open competitive environment. It should be noted that the problem and the solution method presented in this paper are relevant to the e-business context. Moreover, even though our reference service architecture will be the one speci®ed by the telecommunications information networking architecture consortium (TINA-C) [10±12] the presented practices can be applied to other models as well. The approach in this paper is the following. The starting point (Section 2) is the general description of the retailer selection concept, through the presentation of a relevant, overall business case. An outcome of the description of Section 2 is the splitting of the (overall) retailer selection problem into two sub-problems, namely, that of evaluating the quality of a retailer offer and that of determining (con®ning) the set of candidate retailers. Section 3 presents a concise de®nition, mathematical formulation and solution to the problem of evaluating the quality of an offer. Section 4 introduces the learning mechanisms that determine (con®ne) the set of candidate retailers that should be involved in the selection procedure. Section 5 provides a set of indicative results. Finally, Section 6 gives future plans and some concluding remarks. 2. General presentation of the retailer selection concept This section starts from the description of the business

case, through which the role (and importance) of the retailer selection concept can be understood. Section 2.1 provides the description in terms of business level entities (i.e. users and retailers), while in Section 2.2 the description is re®ned by introducing the role of the computational level components. In Section 2.3 the retailer selection problem is analysed and decomposed into sub-problems that are solved in subsequent Sections 3 and 4. 2.1. Description in terms of business entities Assume that a user, wishing to access a speci®c service, can be served by various candidate retailers (CRs), as depicted in Fig. 2. The choice of the most appropriate retailer requires the realisation of the three general phases depicted in Fig. 3. The ®rst general phase involves service independent features like user authentication, authorisation, etc. It involves the user and an entity that will be called default retailer (DR). In essence, at the end of this phase the user is enabled to request services. This phase will not be further addressed in this paper. Apart from its role in the ®rst phase, the DR is seen as an entity that is specialised in the assistance of the user in the open competitive communication environment. The DR can accomplish this by providing, maintaining and hosting (essential parts of) the software that will conduct the retailer selection. In this respect, the DR is assumed to play a coordinating role in the second general phase, which is the core of the retailer selection. At this point, the user has expressed the wish to access a given service. Involved in

Fig. 2. A user is found in an area from which he/she wishes to access a given service through the most appropriate retailer. The area falls into the domain of various candidate retailers.

664

M. Louta et al. / Computer Communications 25 (2002) 662±675

Fig. 3. Interactions among the business level entities during the best, candidate retailer selection business case.

this phase will be the user, the DR and the candidate retailers. In general, the set of candidate retailers can be determined by means of a brokerage (a simple directory) service. Nevertheless, some retailers can be eliminated, as will be explained later in this paper (see Section 4). Moreover, there can be retailers that choose not to make offers to users. In general, retailer selection is founded on general and service speci®c user preferences and retailer policies. In the third phase of the business case the result of the selection is available, and hence an association between the user and the selected retailer can be established and the service usage can possibly start. At this point, some concepts that are complementary (and can be readily integrated) to the business case can be outlined. Speci®cally, the business case can be extended to assist retailers in accounting for their interests. Retailers make offers to users. Additionally, they can be provided with information on the agreements they fail to establish

(e.g. explicitly, through the result and information distribution procedure that is indicated in Figs. 3 and 5 on the last bullet of the retailer selection phase). This information can be exploited for determining whether there should be some modi®cation in the retailer policies (e.g. price reduction, alteration of set of quality levels offered, negotiation strategy modi®cation, etc. [25±29]). Moreover, it should be noted that an alternate (but not completely different) version of the business case could incorporate an auction model. In that case the overall retailer selection concept would consist of an information exchange sub-phase, the auction realisation (this differentiates the business case version) and ®nally the selection. 2.2. Description in terms of computational level concepts A TINA-compliant computational level model of the business case is depicted in Fig. 4. Of interest to our study

Fig. 4. TINA-like computational model for the best candidate retailer selection business case.

M. Louta et al. / Computer Communications 25 (2002) 662±675

665

Fig. 5. Interactions among the computational components in the context of the best, candidate retailer selection business case.

is the TINA access session concept. In general, a session is de®ned as the temporary relationship among a group of objects that are assigned to collectively ful®l a task for a period of time. The TINA access session is a service independent concept, which can be seen as the gateway to any speci®c service usage. It comprises activities that allow user authentication, user pro®le control (inspection), and service invocation. The Initial Agent (IA) is the component that enables the initial access to a domain. The User Agent (UA) component represents the user beyond the terminal e.g. in the default retailer domain. Its role is to intercept and process user requests. The UA has subordinate objects (SOs) that maintain user speci®c information. For brevity, these objects are not shown in full detail in Fig. 5. The UA and SO components maintain user pro®le related information e.g. preferences, requirements and constraints regarding certain services, service subscriptions, etc. The User Application (UAP) models the entity (user interface) with which the

user is confronted in access session mode. The Provider Agent (PA) represents the retailer in the user domain. The UA invokes the Service Factory (SF) for initiating a service. The overall retailer selection task requires a computational component that will act on behalf of the user. Its role will be to capture the user preferences, requirements and constraints regarding the requested service, to deliver them in a suitable form to the appropriate retailer entity, to acquire and evaluate the corresponding retailer offers, and ultimately, to select the most appropriate retailer. As a second step, retailer selection requires an entity that will act on behalf of each candidate retailer. Its role would be to collect the user preferences, requirements and constraints and to make a corresponding offer, taking also into account the underlying connectivity providers. The following key extensions are made so as to cover the functionality that was identi®ed above. First, the UA is extended, by being assigned with the role of selecting on behalf of the user the best retailer. Second, the Retailer Agent (RA) is introduced and assigned the role of promoting the services offered by a candidate retailer. In other words, the UA possesses the user preferences, requirements and constraints from a pro®le, interacts with the RAs of the candidate retailers so as to obtain their offers, and selects the most appropriate retailer for the provision of the desired service. The RA promotes the offers of a candidate retailer, interacts with UAs, and the underlying connectivity provider mechanisms. Fig. 5 presents in more detail the interactions among the computational level components. The detailed description of these interactions is omitted for brevity. 2.3. General analysis of the retailer selection problem

Fig. 6. Approach for (phases in) the solution of the retailer selection problem.

This section analyses the best candidate retailer selection problem that occurs in the second phase of the business case. A general version of the problem can be described as follows. Given (a) a user wishing to access a certain service,

666

M. Louta et al. / Computer Communications 25 (2002) 662±675

possible version of) the problem of evaluating the quality of the retailer offer, which should be solved in the context of the retailer selection phase. Through this work it is shown that the problems can be reduced to well known optimisation problems, which can be solved by relevant standard algorithms. The second area is the presentation of learning mechanisms for reducing the set of candidate retailers, and hence, the associated computations, interactions and resource consumption required. Moreover, and before the above areas, a business case is de®ned through which the role of the best candidate-retailer selection problem is explained. 3. Evaluating the quality of the offer of a retailer

Fig. 7. User-u wishes to access service-s, which is composed of different service features.

(b) user preferences, requirements and constraints regarding the features of the service, and (c) a set of candidate retailers and their offers (e.g. cost at which each service feature Ð quality level combination is provided), ®nd the retailer that best matches a service quality and cost related criterion. The problem above can be analysed as depicted in Fig. 6. In general, the core of the selection process requires a method for evaluating the quality of each retailer offer. This paper (Section 3) includes a mathematical description of a general version of the problem. At a next step the problem version is formulated as a 0±1 linear programming problem [30,31], and a brief outline of computationally ef®cient solution algorithms is presented. The formulation as a linear problem is one of the possible approaches that can be followed. Linear models are an ef®cient compromise to the trade-off of having the essential features of a problem, while maintaining the simplicity of the relations. Refs. [32,33] are recent work samples that have followed a similar approach. Regarding the determination of candidate retailers, the ®rst and obvious approach engages all the possible candidate retailers in the negotiation. The second one, which is considered in Section 4 of this paper, is motivated by the fact that extensive negotiation can possibly entail heavy computations and numerous interactions. In this perspective, learning from experience [34±36] can be exploited, to con®ne the set of candidate retailers. This aspect will be based, in this paper, on retailer rating mechanisms, which take into account the quality of previous retailer offers (performance criterion) and their past performance in consistently satisfying expectations (reliability criterion). The contribution of this paper lies in the following areas. First, the de®nition and mathematical formulation of (one

In this section, we describe in more detail a version of the logic that underlies the UA and RA interactions. As already presented the UA interacts with the RA of each candidate retailer r [ R; where R denotes the overall set of candidate retailers. The aims of the UA±RA interactions are the following. First, to supply the RA with user preferences and constraints regarding the speci®c service. Second, to obtain the corresponding retailer offers. Third, to select the retailer that makes the best offer. The tasks outlined require a method that will enable the assessment of the quality of the offer of each retailer. In this respect, Sections 3.1±3.3 include the de®nition, formulation and solution, respectively, of a problem that can be used for evaluating the quality of the offer of a candidate retailer r: Based on this problem, Section 3.4 describes the resulting retailer selection algorithm and, in ®ner detail, the functionality of the involved computational level components (namely, UA and RA). 3.1. Formal problem statement Each UA acts on behalf of a user u, whose pro®le is known. User u wishes to use a given service s. A fundamental assumption at this point is that service s is composed of a set of distinct service features (e.g. see Fig. 7), which will be denoted as SF(s). Furthermore, let us assume without loss of generality that these service features are offered (supported) by all candidate retailers. This assumption can readily be relaxed as will be explained in subsequent sections (e.g. see in Section 5 the discussion referring to Table 2). Among these service features, of interest to the user are those designated in the user pro®le and will be denoted as SF…u; s† …SF…u; s† # SF…s††: Each service feature i [ SF…s† has an associated set of possible quality levels, represented by the set Q(i). A quality level can be seen as the speci®cation of the (perhaps range of) values of quality parameters that are relevant to the service feature. The set of quality levels that are in line with the user pro®le is denoted by Q…u; i† (i [ SF…u; s†). It holds that Q…u; i† # Q…i†: The user preferences and the retailer policies determine each of these quality level sets.

M. Louta et al. / Computer Communications 25 (2002) 662±675

The anticipated user satisfaction level (measure) that results from the assignment of service feature-i at quality level-i is denoted as bSQ …i; j† (…i [ SF…u; s†; j [ Q…u; i††: Some clari®cations are necessary regarding these parameters. In practice, it is not necessary (or even expected) that the users will be the entities that explicitly con®gure these values (even though this cannot be excluded in certain cases). A realistic assumption is that the DR, being in charge of assisting the user in the open competitive environment, has a solid interest in as accurately as possible re¯ecting the user views in these parameters. In this respect, the DR can be the entity that con®gures the values based on the service feature characteristics, the user preferences and requirements, and by co-operating with retailers and service providers for exploiting their experience. The method for determining the appropriate values can rely on experiments, user trials and the experience obtained during the service provision. The provision of examples on the determination of the bSQ …i; j† values for various services and user classes is a standalone feature left for a future version of this study. The associated price (tariff) that will be imposed on the user by retailer r for the assignment of service feature-i at quality level-i is denoted as pSQ …r; i; j† …r [ R; i [ SF…u; s†; j [ Q…u; i††: The objective of our problem is to ®nd a service con®guration pattern, i.e. an assignment ASQ …r† of service features i…i [ SF…u; s†† to quality levels j…j [ Q…u; i††; that is optimal for retailer r: The assignment should maximise an objective function f …r; ASQ …r†† that models the quality of the retailer r offer. Among the terms of this function there can be the overall anticipated user satisfaction level that results from the assignment, which is expressed by the function b…ASQ …r††; and the price (tariff) at which retailer r will provide the assignment, which is expressed by the function p…r; ASQ …r††: Of course, one of the two factors (anticipated user satisfaction or price of the assignment) can be omitted in certain variants of the general problem version considered in this paper. The constraints of our problem are the following. First, each service feature-i…i [ SF…u; s†† should be assigned to only one quality level-j…j [ Q…u; i††: Second, a cost-related constraint can be imposed. As an example, a value pmax can be de®ned for representing the maximum price (tariff) that can be afforded by the user for the service usage. The pmax value can be seen as an expression of the user constraints. The corresponding mathematical description of the constraint is p…r; ASQ …r†† # pmax : The third problem constraint refers to the anticipated user satisfaction level (measure), which should not be lower than a given value Bmin (this may be seen as an expression of the user requirements). The corresponding mathematical description of the constraint is b…ASQ …r†† $ Bmin : As an example the model of Fig. 7 can be considered. A given user wishes to access a service. The service consists of four service features, each offered at three quality levels. The user pro®le indicates that the user is interested in 3

667

out of 4 service features. Moreover, these service features may be offered to the user in only 2 of the 3 allowable quality levels. A bene®t (measure of the anticipated user satisfaction) will be derived from the provision of a service feature at an associated (allowable) quality level. Thus, we may observe the following: (i) {sf1 ¼sf4 }; (ii) SF…u; s† ˆ {sf2 ; sf3 ; sf4 }; (iii) Q…u; sf2 † ˆ {q22 ; q23 }; etc. The overall problem can be formally stated as follows: Problem 1: [Evaluation of the Quality of the Retailer-r Offer]. Given (a) a user u who wants to use a service s, (b) the pro®le of user-u, (c) the set of service features SF(u,s) of service s that are of interest (relevant) to user u (this set is formed by the service speci®cation, the user pro®le and the retailer capabilities), (d) the set of quality levels Q…u; i† at which each service feature i…i [ SF…u; s†† can be offered, according to the service speci®cation, the retailer capabilities and the preferences of user u, (e) the anticipated user satisfaction level bSQ …i; j† (expressing the user preferences), which derives from the assignment of service feature i…i [ SF…u; s†† to quality level j…j [ Q…u; i†† (f) the price pSQ …r; i; j† that retailer r associates with the assignment of service feature i…i [ SF…u; s†† to quality level j…j [ Q…u; i††; (g) the upper bound on the overall price (tariff) pmax that the user can afford for the service usage (this value is an expression of the user constraints), (h) the lower bound Bmin on the anticipated user satisfaction level that has to be experienced during the service usage ®nd the best service con®guration pattern, i.e. assignment of service features to quality levels ASQ …r†; that optimises an objective function f …r; ASQ …r†† that is related to the overall anticipated user satisfaction b…ASQ …r†† and price p…r; ASQ …r†† suggested by the assignment, under the constraints p…r; ASQ …r†† # pmax ; b…ASQ …r† # Bmin and that each service feature is assigned to exactly one quality level. The above general problem version is open to various solution methods. Its generality partly lies in the fact that the objective and the constraint functions are open to alternate implementations. The problem statement can be distinguished from the speci®c solution approach adopted in Section 3.2. 3.2. Optimal formulation In this section, the problem above is formulated as a 0±1 linear programming problem [30,31]. The experimentation and comparison with important alternate formulation approaches is a stand-alone issue for future study. In order to describe the assignment ASQ …r† of service features to

668

M. Louta et al. / Computer Communications 25 (2002) 662±675

quality levels, the decision variables xE…i; j† …i [ SF…u; s†; j [ Q…u; i†; which take the value 1(0) depending on whether the service feature-i is (is not) assigned to quality level-j; are introduced. The problem of obtaining the most appropriate assignment ASQ …r† may be obtained by reduction to the following optimisation problem. Problem 1: [Evaluation of the quality of the retailer-r offer]. Maximise: X X f …r; ASQ …r†† ˆ ‰cB bSQ …i; j† i[SF…u;s† j[Q…u;i†

2 cP pSQ …r; i; j†ŠxSQ …i; j† subject to X xSQ …i; j† ˆ 1 ;i [ SF…u; s†

…1†

…2†

j[Q…i†

b…ASQ …r†† ˆ

X

X

i[SF…u;s† j[Q…u;i†

p…r; ASQ …r†† ˆ

X

bSQ …i; j†xSQ …i; j† $ Bmin

X

i[SF…u;s† j[Q…u;i†

…3†

pSQ …r; i; j†xSQ …i; j† # pmax …4†

ASQ …r† ˆ {xSQ …i; j†ui [ SF…u; s†; j [ Q…u; i†}

…5†

Relation (1) expresses the objective of ®nding the best assignment of service features to quality levels that maximises the cost function, which is associated with the overall anticipated user satisfaction and the corresponding price. In other words, Relation (1) expresses the quality of the retailer r offer (or equivalently, the objective function value that is scored by retailer r). Weights cB and cP provide the relative value of the anticipated user satisfaction related part and the price related part. Constraints (2) guarantee that each service feature will be assigned to exactly one quality level. Constraint (3) guarantees that the level of user satisfaction will not be lower than a pre-de®ned value that is dictated by the user requirements. In the same manner, Constraint (4) guarantees that total cost will not exceed a prede®ned value. 3.3. Computationally ef®cient solutions In general, the solution of a linear problem can be a computationally intensive task. However, in case the size of the problem instance is not prohibitively large (as often encountered in this paper), a solution method can be to exhaustively search the solution space. The complexity of Q the search in this case is i[SF…u;s† uQ…u; i†u; i.e. a function of the service features that are relevant to the user and the quality levels at which these service features may be offered. In case the solution space is large the design of computationally ef®cient algorithms that can provide good (nearoptimal) solutions in reasonable time is required. Classical

methods in this respect are simulated annealing [37,38], taboo search [39,40], genetic algorithms [41±44], greedy algorithms [31], etc. Hybrid or user de®ned heuristic techniques may also be devised. 3.4. Retailer selection algorithm Ð role of computational components (UA and RA) Based on the discussion so far this section describes the algorithm, on which the UAs and the RAs base the accomplishment of their tasks. The algorithm should be seen as a more detailed description of the tasks in the retailer selection procedure, also taking into account the mathematical problem in Sections 3.1±3.3. Step 1 The UA component is acquainted with the preferences, requirements and constraints of user u regarding service s. These are expressed by the following data. First, the set of the service features SF(u,s) that are of interest (relevant) to the user. Second, for each service feature i…i [ SF…u; s†† the corresponding set of allowable quality levels Q…u; i†: Third, the values bSQ …i; j† that describe the anticipated user satisfaction level stemming from the provision of service feature-i at quality levelj…j [ Q…u; i††: Fourth, the upper limit on the price pmax and the lower limit on the anticipated user satisfaction Bmin that the user can afford, or wants to experience, respectively, during the service usage. Step 2 The UA obtains the list of candidate retailers, R, and the references of the respective RAs. Step 3 The UA component activates the appropriate negotiator entities (e.g.) threads or mobile agents as speci®ed in Ref. [3]). Each negotiator entity will undertake the interactions with a candidate retailer r [ R: The negotiator entities will be under the control of the UA. Step 4 Each negotiator entity obtains the offer of a retailer r [ R for the user preferences, requirements and constraints regarding service-s. These are expressed by the prices pSQ …r; i; j† associated with the provision of service feature-i at quality level-j. Step 5 Each negotiator entity evaluates the quality of the retailer offer by solving the appropriate instance of problem 1. The result (if feasible) is sent to the UA. Step 6 The UA selects a retailer by comparing the objective function values that each retailer has scored. Step 7 End.

4. Determining (con®ning) the set of candidate retailers This section describes the method for con®ning the set of candidate retailers, so as to reduce the required interactions and the associated computations and resource consumption. In other words, this section provides enhancements to the UA intelligence, by incorporating learning-from-experience

M. Louta et al. / Computer Communications 25 (2002) 662±675

concepts. Learning refers to a component's ability to use the information it has obtained from the environment, in order to improve (modify) its decisions and behaviour. As already stated the reduction of the set of candidate retailers will be based on, so-called, retailer rating mechanisms. The rationale of these mechanisms is presented in Section 4.1. Section 4.2 provides the mathematical framework that underlies the rating mechanisms. Section 4.3 presents the revised version of the UA and RA intelligence. 4.1. Retailer rating fundamentals The UA can decide to con®ne the set of candidate retailers based on an estimation of the retailers' expected behaviour. In our approach, this estimation comprises two factors. The ®rst is a measure of the quality of the previous offers that have been made by the retailer, and is called performance criterion. The second is the reliability criterion. Its aim is to re¯ect whether the service ®nally provided to the user corresponded to the agreement reached during the negotiation phase. Our approach is further analysed in the following paragraphs. The performance criterion is motivated by the fact that there may be different levels of user satisfaction with respect to the various retailers' offers. In this respect, there may be retailers that, in principle, do not satisfy the user with their offer. Hence, recording the previous experience can easily assist the UA in deciding whether or not to negotiate with a speci®c retailer. The reliability criterion covers cases in which the retailer does not honour the agreement (or in other words, does not meet up to the expectations) established in the negotiation phase. In this sense, the reliability criterion introduces the ¯avour of trust between the user and the retailer. Obviously, recording pertinent information may encourage or discourage a UA in negotiating with particular RAs. This part of our work is in¯uenced by notions appearing in Refs. [45,46]. Naturally, the UA should apply the mechanisms for con®ning the set of candidate retailers, in cases it is highly likely that the information on which it will be based is accurate. For example, in environments in which the retailers do not change their policies, as considered in this paper, a learning period is required for the UA to obtain the fundamental information for all the retailers. Likewise, UA updating mechanisms are required in case the retailers will be changing their policies, in order to adapt to the market demand. To this end, several approaches can be found in the literature, e.g. the Boltzman exploration strategy. Moreover, according to a straightforward approach that is adopted in this paper, it can be envisaged that the UA will con®ne the set of candidate retailers, in case criteria, indicating that the essential (fundamental) information is not outdated, are satis®ed. In the opposite case, i.e. when the information available to the UA is likely to be outdated (inaccurate), the negotiation will involve all the candidate retailers.

669

At this point it should be noted that the criteria presented above can be combined, or one of them can be disregarded, without affecting the overall framework envisaged in this section. A more detailed experimentation with alternate (perhaps composite) rating criteria is left for a future version of this study. 4.2. Mathematical description of the retailers rating mechanisms This section provides the formulas that realise the retailer rating mechanisms in the UA. 4.2.1. Formulations for the performance criterion of the retailer rating mechanisms Each retailer r may be rated according to the performance criterion through the following formula: RPpost …r† ˆ RPpre …r† 1 kp …rp…r† 2 E‰rp…r†Š†

…6†

where RPpost …r† and RPpre …r† are the retailer-r performancebased rating after and before the updating procedure, rp…r† is a (reward) function that describes the quality of the retailer-r current offer (with respect to the other retailers), and E‰rp…r†Š is the mean (expected) value of the rp…r† value. In general, the larger the rp…r† value the better the quality of the current offer, and therefore, the more positive the in¯uence on the rating of the retailer. Factor …kp …kp $ 1† determines the relative signi®cance of the new outcome with respect to the old one. In essence, this value determines the memory of the system. Small kp value means that the memory of the system is large. Therefore, good offers will gradually improve the retailer's rating position. It should be noted that a deterioration of the quality of the offer of retailer r, with respect to that made by other retailers, leads to a decreased post rating value, since then the …rp…r† 2 E‰rp…r†Š† quantity is negative. The rp…r† function may be implemented in several ways. In the result sections of this paper, it was assumed without loss of generality that the rp…r† values vary from 0.1 to 1. 4.2.2. Reliability related rating mechanism-update formulae This section introduces the formulas used for the rating of retailers according to the reliability criterion. In general, the approach is similar to that of the previous sections. Each retailer r may be rated according to the reliability criterion through the following formula: RRpost …r† ˆ RRpre …r† 1 kr …rr…r† 2 E‰rr…r†Š†

…7†

where RRpost …r† and RRpre …r† are the retailer-r reliabilitybased rating before and after the updating procedure, rr…r† is a (reward) function re¯ecting whether the service quality is compliant with the picture established during the negotiation phase, and E‰rr…r†Š is the mean (expected) value of the rr…r† value. The kr factor plays the same role as in the performance rating case. It should be noted that the reliability value of the selected

670

M. Louta et al. / Computer Communications 25 (2002) 662±675

Table 1 First set of experiments: (a) description of service preferences in the pro®les of the 10 user classes. The ticked boxes indicate that the users of the class are interested for the corresponding quality level at which the service feature can be provided; (b) description of the retailer policies. Prices at which each service feature-quality level combination is provided. Empty boxes indicate that the corresponding service feature-quality level combination is not supported

(a) User class k1 k2 k3 k4 k5 k6 k7 k8 k9 k10 (b) Retailer R1 R2 R3 R4 R5 R6 R7 R8 R9 R10

QA1

QA2

QA3

u

u u

u

u u

u u

u

QV1

QV2

u

u u u u u

u u

QV3

u a

u a

u u u

u

u

u

u

u

u

u

3

5

2 2 1.8 2 1.8 2.5 1.8 2 2 1.8

3 2.8 2.5 2.8 2.5 2.8

5.1

2.3

4.1

u

u

u

1 0.9 1 1 0.9 0.9 0.8 1 0.5 0.8

1.5 1.2 1.4 1.4 1.2 1.7 1.5

2.5 2.5 2.8 2.5

4.5

QV4

u

u

u u u

1.5

QA4

5 4

7

4 6.9

2.3

retailer is updated after the user ®nally accesses the service. Moreover, this rating requires a mechanism for evaluating whether the service quality was compliant with the picture promised during the negotiation phase.

the estimated retailer rating values formed according to the formulas presented in the previous section. Step 2 The UA obtains the list of candidate retailers, R, and the references of the RAs. Step 3 The UA forms the con®ned set of candidate retailers, Rc …Rc # R†; based on the rating mechanisms presented previously in this section, in case the pertinent fundamental information, on the usual quality of the retailers' offers and their reliability, is likely not to be outdated. In the opposite case, i.e. when the information available to the UA is likely to be outdated, the negotiation will involve all the candidate retailers, i.e. then Rc ˆ R: Step 4 The UA activates the appropriate negotiator entities for undertaking the interactions with the retailers in the con®ned set of candidate retailers, Rc. Step 5 Each negotiator entity obtains an offer from retailer r…r [ Rc † for the user preferences, requirements and constraints regarding service-s: These are expressed by the prices pSQ …r; i; j† associated with the provision of service feature-i at quality level-j. Step 6 Each negotiator entity evaluates the quality of the offer of retailer r…r [ Rc † by solving the appropriate instance of problem 1. The result (if feasible) is sent to the UA. Step 7 The UA selects a retailer by comparing the objective function values that each retailer has scored. Step 8 The retailers' rating values (performance and reliability related) are updated on the basis of Eqs. (6) and (7) respectively. Step 9 End.

5. Results 4.3. Revised retailer selection algorithm and role of computational components (UA and RA) In this section, we describe the algorithm, on which the SUAs and the RAs base the accomplishment of their tasks. The algorithm should be seen as a more detailed description of their sub-tasks involved, taking also into account the schemes of this section. Step 1 The UA is acquainted with the preferences, requirements and constraints of user u regarding service s. These are expressed by the following data. First, the set of the service features SF(u,s) that are of interest (relevant) to the user. Second, for each service feature i…i [ SF…u; s†† the corresponding set of allowable quality levels Q…u; i†: Third, the values bSQ …i; j† that describe the bene®t (user satisfaction) stemming from the provision of service feature-i at quality level-j…j [ Q…u; i††: Fourth, the upper limit on the price pmax and the lower limit on the user satisfaction Bmin that the user may afford, or wants to experience, respectively, during the service usage. Fifth,

The results of this section aim at discussing on the ef®ciency of the overall retailer selection scheme and the learning mechanisms that con®ne the set of candidate retailers. The ef®ciency of the overall retailer selection scheme will be measured with respect to a random retailer selection scheme. Two sets of tests will be used for discussing on these aspects. This section assumes the existence of an area that falls into the domains of R candidate retailers. Users access the area in order to initiate a service usage. In the context of our experiments, it is assumed that users request a videoconference service. A simple and wellknown service has been chosen in order to explain the proposed scheme. Nevertheless, any other service could have been chosen instead. The videoconference service comprises two service features, namely audio and video. In the context of our study, four quality levels have been considered for these service features. Speci®cally, the quality levels that have been de®ned for audio are 8, 16, 32 and 64 kbits/s, respectively. In a similar manner,

M. Louta et al. / Computer Communications 25 (2002) 662±675 Table 2 The set of inappropriate candidate retailers per user class User class

Inappropriate retailers

k1 k2 k3 k4 k5 k6 k7 k8 k9 k10

± 7, 8, ± 7, 9 2, 5, 3, 5, 2, 3, 3, 5, 3, 5, ±

9, 10 7, 8, 9, 10 7, 8, 9, 10 4, 5, 7, 8, 9, 10 7, 9, 10 7, 9, 10

the de®ned quality levels for video are 15, 20, 25 and 30 frames/s, respectively. Regarding the different users that access the area, it is assumed that k user classes exist. In the de®nition of these user classes, we have also assumed that all users in these classes are interested for both service features. However, each user class is interested in different quality levels of these service features. Concerning the implementation issues of our experiment, the whole TINA access session has been implemented in Java [47] in the context of [3]. The OrbixWeb CORBA compliant platform [48] was used for the inter-component communication. Moreover, the UA and the RA have been implemented as intelligent, mobile agents based on the use of the Voyager platform [49]. The pro®les of each user class in the ®rst experiment are presented in Table 1(a). It has been assumed that k ˆ 10: More speci®cally, this table indicates the quality levels for audio and video, i.e. QAj and QVj …1 # j # 4†; respectively, which are of interest to each user class. Regarding the user satisfaction level, the simplest possible assumption has been made in this experiment. Speci®cally, it has been assumed that the users in each class are equally attracted by all the quality levels at which a service feature can be offered. This Table 3 First set of experiments: outcome of the best candidate retailer selection scheme. Objective function value, selected retailer, and improvement with respect to the random retailer selection scheme, per user class involved in the experiment User class Objective function value k1 k2 k3 k4 k5 k6 k7 k8 k9 k10

97.5 96.3 97.5 96.9 95 94.8 93 95.1 95.1 97.5

Retailer selected

Improvement with respect to random retailer selection scheme (%)

9 5 9 10 3 2 6 2 2 9

25.04 27.33 19.49 16.92 29.78 19.38 4.44 17 9.26 40.84

671

assumption will be changed in the second experiment. Nevertheless, in the ®rst experiment it enables the acquisition of an initial set of indicative results that show the behaviour of our schemes. In the light of the assumption made, the problem is reduced to the selection, for each service feature, of the acceptable quality level that will minimally impact the price. Usually, this is the lowest quality level at which the service feature can be provided, according to the user preferences. Table 1(b) describes the retailer policies regarding the features (audio and video) of the videoconference service and the respective quality levels. It has been assumed that uRu ˆ 10 (i.e. 10 retailers are involved in the experiment). More speci®cally, this table indicates the offered quality levels for audio and video for each retailer, as well as the price (expressed in arbitrary values) which is associated with the provision of a service feature at a given quality level. It can be observed that the price at which each retailer offers a service feature Ð quality level combination increases, as the quality level increases. Another aspect that should be noted is that, in order to make the test case more realistic (or general), all retailers do not offer all possible quality levels. Retailers whose maximum offered quality level, regarding a service feature, is below the minimum acceptable by a user class u; constitute the I(u) set, which comprises retailers that are inappropriate for the user class. As an example, user class k2 requires the video component at least at quality level QV2 : In this respect, retailer R7 that offers the video component at quality level QV1 maximum is inappropriate for user class k2. Table 2 presents the set of inappropriate retailers for each user class, taking into account the pro®les and retailer policies described in Table 1. As previously mentioned, the objective of our experiment is to provide indicative evidence of the overall retailer selection scheme, with respect to a random retailer selection scheme. In this respect, Table 3 presents the outcome of the application of the retailer selection scheme for the user classes and retailer policies described in Table 1. Speci®cally, for each user class, the derived value of objective function (1), the selected retailer, and the decrease with respect to the random retailer selection scheme are shown. It should be noted that the bSQ …i; j† coef®cients have arbitrarily been set to be equal to 50. Moreover, with reference to Table 3, from the derived results it can be deduced that the desired service features are always offered at the lowest allowable quality level. This is justi®ed since (a) the users are equally attracted by the different quality levels and (b) the considered allocation yields minimum cost. In general, from the results of Table 3 it is observed that the best candidate retailer-selection scheme exhibits a better performance with respect to the random retailer selection scheme, which on the average is in the order of 20%. This decrease is due to the selection of the most suitable retailer taking into account the user preferences and the retailer policies.

672

M. Louta et al. / Computer Communications 25 (2002) 662±675

Table 4 First set of experiments: (a) performance-based and (b) reliability-based retailer rating, prior to the UA's decision on the con®ned set of candidate retailers User class

R1

R2

(a) Performance ratings k1 7 6 k2 5 3 7 6 k3 k4 8 6 k5 4 3 1 k6 k7 2 k8 4 1 k9 5 1 7 6 k10 (b) Reliability ratings 2 8 k1 k2 2 5 k3 2 4 2 3 k4 k5 1 k6 1 3 k7 1 1 2 k8 k9 1 2 k10 2 4

R3

R4

R5

R6

R7

R8

R9

R10

5 2 5 4 1

8 4 8 7 2 2

4 1 4 3

10 6 10 5 3 4 1 5 4 10

2

9

1

3

3

9 2

1

2 1

2

3 3 9

1

3

6 4 6 5 3 2 2 3 3 7

3

9

4

7

3

10 8

5

7 6

5

8

5 5 3 8 4 2

6

2 2 8 10 6 9 7 4 4 4 4 9

4 1 1 1 1

1

3

5 5 10

As previously mentioned, another objective of our experiment is to provide indicative evidence on the ef®ciency of the learning mechanisms that con®ne the set of candidate retailers. In this respect, Table 4(a) and (b) present the outcome of the application of the retailer rating mechanism

according to the performance and reliability criteria. The results of the tables are explained in the following paragraphs. The retailer ratings have been created by the UA during a learning period and are based on the policies of Table 1 and on Relations (5) and (6). It should be noted that inappropriate retailers are not considered when deciding for the most promising retailers during the con®guration of the con®ned sets. In Table 4(a), the retailers are ranked with respect to the performance criterion, which re¯ects the usual quality of their offer for the various user classes. In this respect, for example, for user class k1 the best candidate retailer (according to the policies of Table 1 and problem 1), R9, is ranked ®rst, while retailers R7 and R10 are ranked second and third, respectively. This can be explained as follows. The users of the class are equally attracted by quality levels QA1 and QA2 for audio, and QV1 and QV2 for video (as can be observed in Table 1). The cheapest combination is …QA1 ; QV1 † as priced by R9. Likewise, the …QA1 ; QV1 † combination as provided by R7 is the second cheapest one. Hence, the rating mechanisms will favour these retailers, and they will be found with a high ranking. In Table 4(b), the retailers are ranked with respect to the reliability criterion, which refers to whether the retailer usually meets the quality expectations raised (or promised) during previous negotiation phase. In order to test this aspect each retailer has been associated with a reliability probability, i.e. a measure of the likelihood that the retailer delivers a service that is compliant with the agreement established in the negotiation phase. This probability has been set to 0.9 for R5, 0.8 for R1, 0.7 for R7, 0.6 for R2 and

Table 5 First set of experiments. Decision on the con®ned set of candidate retailers, and outcome of the retailer selection scheme (i.e. objective function value and selected retailer), for each user class, on the basis of the performance rating and reliability rating User class

Con®ned set of retailers

Objective function value

Retailer selected

Improvement (%)

Performance ratings k1 k2 k3 k4 k5 k6 k7 k8 k9 k10

7, 2, 7, 5, 3, 1, 1, 2, 2, 7,

9, 10 3, 5 9, 10 8, 10 4, 6 2, 4 6 4, 8 4, 8 9, 10

97.5 96.3 97.5 96.9 95 94.8 93 95.1 95.1 97.5

9 5 9 10 3 2 6 2 2 9

25.04 27.33 19.49 16.92 29.78 19.38 4.44 17 9.26 40.84

Reliability ratings k1 k2 k3 k4 k5 k6 k7 k8 k9 k10

1, 1, 1, 1, 1, 1, 1, 1, 1, 1,

5, 7 3, 5 5, 7 5, 2 3, 6 2, 6 6 2, 6 2, 6 5, 7

97.4 96.3 97.3 96.6 95 94.8 93 95.1 95.1 97.4

7 5 5 5 3 2 6 2 2 7

21.8 27.33 13.04 11.56 29.78 19.38 4.44 17 9.26 38.47

M. Louta et al. / Computer Communications 25 (2002) 662±675 Table 6 Second set of experiments: description of service preferences in the user pro®les of the 10 user classes that are involved in the experiments User class

QA1

QA2

k1 k2 k3 k4 k5 k6 k7 k8 k9 k10

50

55 50

50 50

51 50

50 50 50

QA3

QA4

51.5 50

50.3

53

50.3

51.5

QV1

QV2

50

50.5 50 50.5 50 50

50 51 50

50

53

50

50

51

QV3

QV4

50 53 50

53 51

50 51 51

53

R9, 0.5 for R3, R6 and R10, and 0.2 for R4 and R8. Speci®cally, with probability 0.9 retailer R5 complies with the promises, while retailer R9 maintains its promises with probability 0.6. A mixture of extreme and moderate values has been chosen in order to test the schemes under diverse circumstances. As expected and as can be observed in Table 4(b) the reliability criterion alters the retailer ranking. As an example, the difference in user class k1 should be noted. Speci®cally, according to the performance criterion retailer R9 was ranked ®rst. Nevertheless, when reliability aspects are taken into account, retailer R9 withdraws in fourth place, considering that (according to our arbitrary assumption) the probability of remaining compliant with the offer is 0.6. On the contrary, according to the performance criterion retailer R5 is ranked fourth. However, this retailer is highly likely (probability 0.9) to remain compliant with the offer, and therefore, according to the reliability criterion it is ranked ®rst. Having the ratings of Table 4 as a basis, the con®ned set of candidate retailers is formed. Table 5 present the behaviour of the retailer selection scheme after the end of the learning period. Speci®cally, for each user class, the con®ned set of candidate retailers, the objective function Table 7 Second set of experiments: outcome of the best candidate retailer selection scheme. Objective function value, selected retailer, and improvement with respect to the random retailer selection scheme, per user class involved in the experiment User class Objective function value

k1 k2 k3 k4 k5 k6 k7 k8 k9 k10

102 96.5 97.5 97.3 96.5 94.8 95.7 96.5 96.1 97.9

Retailer selected Improvement with respect to random retailer selection scheme (%) 5 3 9 5 4 2 6 4 2 10

51.02 17.92 13.94 15.07 23.49 16.13 17.28 20.86 9.26 28.09

673

Table 8 Second set of experiments: (a) performance-based, and (b) reliability-based retailer rating, prior to the UA's decision on the con®ned set of candidate retailers User class

R1

R2

(a) Performance ratings k1 7 3 k2 6 4 8 7 k3 k4 7 3 k5 3 3 1 k6 k7 2 k8 4 3 k9 5 1 9 7 k10 (b) Reliability ratings 2 4 k1 k2 2 5 k3 2 4 2 3 k4 k5 1 k6 1 3 k7 1 1 2 k8 k9 1 2 k10 2 5

R3

R4

R5

R6

R7

R8

R9

R10

2 1 5 2 4

5 3 9 5 1 2

1 2 4 1

8 5 10 8 2 4 1 2 4 10

4

10

6

9

3

6 6

1

2 4

6

5 3 2

4

1

7 4 6 4 2 2 2 3 3 7

3

9

5

8

3

9 7

5

7 5

4

5

5 6 3 8 6 3

6

1 2 8 10 6 10 8 4 4 4 4 9

3 1 1 1 1

1

3

5 5 10

value and the selected retailer are shown. In our test case, the three most appropriate retailers were selected to constitute the con®ned set of candidate retailers. It should also be stressed that for user classes k1, k3, k4 and k10 different retailers have been selected when different criteria have been applied. This is due to the introduction of different degrees of reliability for different retailers. This introduction favours the selection of retailers that remain compliant with their offers, with respect to those that can make good offers during the negotiation phase, but can also violate their promise during the service provision phase. For example, for user class k1 retailer R9 is selected according to the performance criterion. However, since retailer R9 has been associated with a high unreliability probability (arbitrarily by our working assumption) its ranking with respect to the reliability criterion declines, and retailers R1, R5 and R7 become the most promising ones. Among these, retailer R7 that makes the best offer is selected. The pro®les and service preferences of the different user classes in the second set of experiments are shown in Table 6. In this test case it is assumed the users of each class exhibit a different attraction volume for each permissible service feature Ð quality level combination. This is re¯ected in the bSQ …i; j† coef®cients. It can be observed that in some cases higher quality levels are more attractive to users. The objective of the retailer selection scheme is to ®nd the assignment of service features to quality levels that maximises the value of objective function (1). The retailer policies shown in Table 1(b) are also valid in this set of experiments.

674

M. Louta et al. / Computer Communications 25 (2002) 662±675

Table 9 Second set of experiments. Decision on the con®ned set of candidate retailers, outcome of the retailer selection scheme (i.e. objective function value and selected retailer), and service con®guration pattern (allocation of service features to quality levels), for each user class, on the basis of the performance rating, and reliability rating User class

Con®ned set of retailers

Performance ratings k1 2, 3, k2 3, 4, k3 7, 9, k4 2, 3, k5 1, 4, k6 1, 2, k7 1, 6 k8 2, 4, k9 2, 4, k10 5, 8, Reliability ratings 1, 5, k1 k2 1, 3, k3 1, 5, k4 1, 2, k5 1, 3, k6 1, 2, k7 1, 6 k8 1, 2, k9 1, 2, k10 1, 5,

5 5 10 5 6 4 6 8 10 7 5 7 5 6 6 6 6 7

Objective function value

Retailer selected

Improvement (%)

Assigned QAi

Assigned QVj

102 96.5 97.5 97.3 96.5 94.8 95.7 96.5 96.1 97.9

5 3 9 5 4 2 6 4 2 10

51.02 17.92 13.94 15.07 23.49 16.13 17.28 20.86 9.26 28.09

QA2 QA3 QA1 QA2 QA3 QA2 QA4 QA3 QA1 QA1

QV1 QV2 QV1 QV2 QV3 QV3 QV2 QV3 QV3 QV2

102 96.5 97.4 97.3 95.1 94.8 95.7 95.1 96.1 97.6

5 3 7 5 6 2 6 6 2 5

51.02 17.92 9.83 15.07 10.85 16.13 17.28 7.542 9.26 24.58

QA2 QA3 QA1 QA2 QA3 QA2 QA4 QA3 QA1 QA2

QV1 QV2 QV1 QV2 QV3 QV3 QV2 QV3 QV3 QV2

Table 7 presents the outcome of the application of the retailer selection scheme for the user classes of Table 6 and retailer policies described in Table 1(b). Speci®cally, for each user class, the derived value of objective function (1), the selected retailer, and the decrease with respect to the random retailer selection scheme are shown. From the results of Table 7 it is observed that the best candidate retailer-selection scheme exhibits a signi®cantly better performance with respect to the random retailer selection scheme. In a manner similar to the ®rst set of experiments, Table 8(a) and (b) present the outcome of the retailer rating mechanism taking into account the performance and reliability criteria, for the user classes described in Table 6 and the retailer policies described in Table 1(b). The retailer ratings have been created by the UA during a learning period. Table 9 present the behaviour of the overall retailer selection procedure (after the end of the learning period). More speci®cally, the results provide for each user class, the con®ned set comprising the most promising candidate retailers for the speci®c service requested, the derived value of objective function (1), the selected retailer and the quality levels at which the service features (audio and video) were provided. Comparing to the corresponding results in the ®rst set of experiments (Table 5), it may be observed that the value of objective function (1) is increased up to 5% approximately, for user class k1. The increase is due to the attraction of users towards higher quality levels that are more expensive. It can also be stressed that for some user classes, e.g. k3,

k5, k8 and k10 different retailers have been selected compared to the ®rst set of experiments. This, again, is due to the introduction of different levels of attraction for different quality levels. This introduction entails the selection of a different retailer, which offers a better balance between user satisfaction and derived cost. 6. Conclusions and future work The highly competitive communications markets of the future should encompass mechanisms for enabling users to ®nd and associate with the most appropriate retailers, i.e. those offering adequate quality services in a cost ef®cient manner. This paper presented such mechanisms. Our starting point was the de®nition of a business case, through which the role of the best candidate-retailer selection problem was explained. In the sequel, the problem was analysed and the identi®ed sub-problems were concisely de®ned, mathematically formulated and solved. The identi®ed components of the best candidate-retailer selection problem were targeted to the evaluation of the quality of a retailer offer and the reduction of the set of candidate retailers by exploiting learning from experience notions. In the ®nal sections, the paper results are provided and concluding remarks are made. Directions for future work include, but are not limited to the following. First, the realisation of further wide scale trials, so as to experiment with the applicability of the framework presented herewith. Second, the experimentation

M. Louta et al. / Computer Communications 25 (2002) 662±675

with alternate approaches for evaluating the quality of the retailers' offers and their reliability. Third, the experimentation with approaches that combine the performance and the reliability criteria.

[20]

References

[22] [23]

[1] H. Ungerer, N. Costello, Telecommunications in Europe: Free Choice for the User in Europe's 1992 Market, Of®ce for Of®cial Publications of the European Communities, Brussels, 1990. [2] H. Zuidweg, M. Campolargo, J. Delgado, A. Mullery (Eds.), Paving the way for an open service market, 6th International Conference on Intelligence in Services and N etworks (IS and N'99) Springer, Berlin, 1999. [3] D. Prevedourou, K. Zygourakis, S. Efremidis, G. Stamoulis, D. Kalopsikakis, A. Kyrikoglou, V. Siris, M. Anagnostou, E. Tzifa, T. Louta, P. Demestichas, N. Liossis, A. Kind, K. Valtari, H. Jormakka, T. Jussila, Use of agent technology in service and retailer selection in a personal mobility context, Comput. Networks J. 31 (1999) 2079± 2098. [4] S. Ghosh, Making business sense of the internet, Harvard Buss. Rev. 76 (2) (1998). [5] R. Kalakota, A. Whinston, Electronic Commerce: A Manager's Guide, Addison-Wesley, Reading, MA, 1997. [6] Electronic commerce, Special Issue Comput. Networks J. 32 (6) (2000). [7] J. Delgado, G. Stamoulis, A. Mullery, D. Prevedourou, K. Start (Eds.), Telecommunications and IT convergence-towards service e-volution, 7th International Conference on Intelligence in Services and Networks (IS and N 2000) Springer, Athens, 2000. [8] S. Trigila, A. Mullery, M. Campolargo, J. Hunt, Service architectures and service creation for integrated broadband communications, Comput. Commun. 18 (11) (1995). [9] S. Trigila, K. Raatikainen, B. Wind, P. Reynolds, Mobility in longterm service architectures and distributed platforms, IEEE Personal Commun. 5 (4) (1998). [10] T. Magedanz, TINA Ð architectural basis for future telecommunications services, Comput. Commun. 20 (4) (1997). [11] W. Barr, T. Boyd, Y. Inoue, The TINA initiative, IEEE Commun. Mag. 31 (3) (1993). [12] F. Dupuy, G. Nilsson, Y. Inoue, The TINA consortium: toward networking telecommunications information services, IEEE Commun. Mag. 33 (3) (1995). [13] A. Lazar, Programming telecommunication networks, IEEE Network 11 (5) (1997). [14] A. Lazar, K. Lim, F. Marconcini, Realising a foundation for programmability of ATM networks with the binding architecture, IEEE J. Selec. Areas Commun. 14 (7) (1996). [15] M. Tag, Service creation environment engineering, Proceedings of Interworking '96 Conference, Nara, Japan 1996. [16] P. Demestichas, F. Westerhuis, M. Guyot, C. Loftus, E. Sherratt, M. Anagnostou, Proceedings of IEEE International Conference on Intelligence Networks (ICIN '98), Bordeux, France 1998. [17] E. Tzifa, P. Demestichas, M. Lambrou, A. Kaltabani, S. Kotrotsos, M. Anagnostou, Advanced service creation environment: practices, implementation and application, in: J.-Y. Roger, B. Stanford-Smith, P.T. Kidd (Eds.), Technologies for the Information Society: Development and Opportunities, IOS Press, Amsterdam, 1998. [18] B. Wind, M. Samarotto, P. Nicosia, M. Lambrou, E. Tzifa, Enhancing the TINA architectural framework for personal mobility support, Proceedings of 5th Conference on Intelligence in Services Networks (IS and N'98) Antwerb, Belgium 1998. [19] S. Vinoski, CORBA: integrating diverse applications within distrib-

[21]

[24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] [47] [48] [49]

675

uted heterogeneous environments, IEEE Commun. Mag. 35 (2) (1997). Object Management Group (OMG) Web site, www.omg.org, January 2001. Mobile software agents for telecommunications. Feature Topic IEEE Commun. Mag. 36 (7) (1998). P. Morreale, Agents on the move, IEEE Spectrosc. 35 (4) (1998). Active and programmable networks, Special Issue IEEE Network (1998). Active networks, Feature Topic IEEE Comput. (1999). P. Dutta, Strategies and Games: Theory and Practice, MIT Press, Cambridge, MA, 1999. R. Lewicki, D. Saunders, J. Minton, Negotiation: Readings Exercises and Cases, McGraw-Hill, Boston, 1999. G. Shell, Bargaining for Advantage: Negotiation Strategies for Reasonable People, Viking, Penguin Books, New York, 1999. J. Morris, P. Maes, Negotiating beyond the bid price, Proceedings of ACM Conference on the Human Factors in Computing Systems (CHI 2000), The Hague, Netherlands 2000. J. Morris, P. Ree, P. Maes, Sardine: dynamic seller strategies in an auction marketplace, Proceedings of the ACM Conference on Electronic Commerce (EC-00), Minnepolis, USA, October 2000 2000. H. Salkin, Integer Programming, Addison-Wesley, Reading, MA, 1975. C. Papadimitriou, K. Steiglitz, Combinatorial Optimization: Algorithms and Complexity, Prentice Hall, Engelwood Cliffs, NJ, 1982. M. Fan, J. Stallaert, A. Whinston, A web-based ®nancial trading system, IEEE Comput. (1999). T. Hamada, S. Hogg, J. Rajahalme, C. Licciardi, L. Kristiansen, P.F. Hansen, Service quality in TINA: quality of service trading in open network architecture, IEEE Commun. Mag. (1998). R. Sutton, A. Barto, Reinforcement Learning: An Introduction (Adaptive Computation And Machine Learning), MIT Press, Cambridge, 1998. R. Michalski, J. Carbonell, T. Mitchell, Machine Learning: An Arti®cial Intelligence Approach, Morgan Kaufmann, Los Altos, CA, 1986. V. Cherkassky, F. Mulier, Learning from Data: Concepts, Theory, and Methods (Adaptive and Learning Systems for Signal Processing, Communications and Control Series), Wiley, New York, 1998. E. Aarts, J. Korts, Simulated Annealing and the Boltzmann Machines, Wiley, New York, 1989. P. van Laarhoven, E. Aarts, Simulated Annealing: Theory and Applications, Kluwer Academic, Dordrecht, 1988. F. Glover, M. Laguna, Taboo Search, Kluwer Academic, Hingham, MA, 1998. F. Glover, E. Taillard, D. de Werra, A guide to Taboo search, Ann. Oper. Res. 41 (1993). D. Goldbgerg, Genetic Algorithms in Search, Optimisation and Machine Learning, Addison-Wesley, Reading, MA, 1989. L. Davis, Handbook of Genetic Algorithms, Van Nostrand Reinhold, New York, 1991. R. Haupt, S. Haupt, Practical Genetic Algorithms, Wiley, New York, 1997. Z. Michalewicz, Generic Algorithms 1 Data Structures ˆ Evolution Programs, Springer, Berlin, 1995. A. Moukas, R. Guttman, G. Zacharias, P. Maes, Agent-mediated electronic commerce: an MIT media laboratory perspective, Int. J. Electron. Comm. (1999). R. Guttman, A. Moukas, P. Maes, Agent-mediated electronic commerce: a survey, Knowl. Engng Rev. (1998). J. Gosling, B. Joy, G. Steele, The Java Language Speci®cation, Addison-Wesley, Reading, MA, 1996. IONA Technologies Web site, www.orbix.com, May 2001. ObjectSpace Inc. Web site, www.objectspace.com, May 2001.