ASPASIA Knowledge Systems - Online Case-Based Reasoning (CBR) WEB-Services in Knowledge Management

Online Case-Based Reasoning (CBR) WEB-Services in Knowledge Management Applications

The initial point for presenting this system are the results of an European project Aspasia has been involved in [1]. Consequently and on the basis of this former RTD actions [2, 3, 4] Aspasia presents in this section an enhancement of a specific type of decision support methodology - a new approach in Case Based Reasoning (CBR) for the general application in composing and execution of WEB-Services. The realisation trait of this system is the use of a closed loop approach considering and taking into account statistically evaluated user and experience feedback information implicitly. This closed loop structure assures a stable dynamically growing of the internal case bases and optimal adap-tation to the user resp. systems needs. Furthermore the description will be given for introducing the CBR methodology as decision making and inferencing unit in web service composing techniques.

Due to the fact that nowadays new knowledge creation and generation is rapidly accelerating, the development of efficient tools and systems for knowledge management and reasoning is essential. A main application field in the wide area of knowledge management is designated to decision sup-port systems - systems which provide comprehen-sive support in gathering, organisation, refine-ment and distribution of knowledge. Within the next sections the main and significant terms and definitions of CBR are given, whereas in the following part the different application fields and realisation areas are touched [5]. Exemplary the approach of a CBR-Wrapper [6] composed for product configuration used in a personal computer sales system will be mentioned. In [8] the new term "utility" [9] is introduced in CBR methodology. Alternatively a further approach of CBR modelled and designed as a closed loop process is pres-ented and illustrated. As a further meanwhile more important application field - the web service technology - is discussed. In this context Aspasia proposes and introduces the CBR approach as the fusion of inferencing and Knowledge Base in web service composing processes.

Case Based Reasoning (CBR)

Generally problem solving related to CBR is based on the main assumption, that "similar problems have similar solutions". Accordingly the general structure of a CBR-System is illustrated in (Fig. 1), where the relation-ship between user oriented task resp. problem description and the recommended solution (knowledge object, artefact) is traditionally based on similarity measures given in a so called "case base". In this case base the problem descriptions and solutions are stored as coupled object pairs.

Hence, case resp. solution / subjects are retrieved from the case base using a similarity relationship. In this sense the decision making process is mainly dependent on the quality of determining a more likely precise similarity measure - some kind of function that assesses this similarity and represent it by a numeric value (similarity coefficient). Generally the case with the highest similarity is selected for recommendation. CBR can be formalised as a four-step process:

Retrieve: Given a target problem, retrieve cases from the case base that are relevant to solving it.
Reuse: Map the solution from the previous case to the target problem, and try to adapt the solution as needed to fit the new one.
Revise: After mapping the previous solution to the target one, the new solution should be evaluated - in case of bad functioning revise it.
Retain: After positive adaptation of the solution to the target problem, store the resulting experience as a new case in the case base.
In contrary to rule based decision making procedures the CBR-approach has its advantages in all fields, where the knowledge about the decision making related facts are not clearly and precise describable and elusive.

Application Fields of CBR

The CBR methodology is often implemented in areas, where some kind of "reuse" is appropriate, like elec-tronic design reuse, reuse of SW packages and, component based SW devel-opment. Therefore applications of the CBR approach can be found in different fields and areas like electronic commerce, planning issues (project management), diagnosis, CBR-supported planning Design - generally spoken in all areas, where reasoning and decision making might be based on existing "similar" solutions and cases. At present one of the most important commercial application is in e-commerce for example the handling of distributed information and sales support for customised e-commerce products.

In [6] a framework for product configuration is introduced. It consists of a configurator and a wrapper for the configurator. The application field hereby is within e-commerce a specific personal computer sales system, whereby the goal of product configuration is to assemble PC products and components by arranging a predefined set of product parts, so that the product satisfy the user requirements by "observing the connectivity of the parts and limitations on the resources necessary for assembling the products" [7]. The CBR wrapper generates a task resp problem description according to the user query. The problem description itself consists of parts description and a requirements description, whereas the wrapper generates only requirements description using similar cases for a given query. And the solver performs the actual task of product configuration based on the problem description.

Utility Oriented Matching in CBR

The authors in [8] propose in their work to "overcome" the traditional view of CBR, namely "similar problems have similar solutions". They want to extend the CBR view to situations in which "we do not have cases that contain pairs of problem descriptions and solutions in the usual sense". However, in [8] the new term "utility" [9] is introduced in CBR methodology, based on the cognition, that "a similarity measure always tries to approximate some form of utility" [8]. Therefore the authors in [8] proposed an utility-oriented case matching by applying a utility reasoning level (Fig. 2). They overcome the assumption that cases consist of a problem description and a solution description - in their approach "only a solution description is required". From this solution description they compute the so-called "utility description", which means some kind of "utility reasoning" has to be performed. Hereby an "utility function" is introduced, reflecting "elementary references". The authors don"t provide any solution which considers the users feedback - or implicitly internal feedback - it seems that this approach is a pure static approach, because the "utility-function" is determined and loaded by the administrator (expert, maintainer) of the CBR system. This utility ori-ented procedure of reasoning is solely a static ap-proach and due to this facts it might be stated as an open loop system, which is in sense of cybernetics not dynamic and not capable to handle and manage user influences and experiences and accordingly not enables an unsupervised stable growing of the case base.

New Approach of Closed-Loop CBR

The Aspasia authors point out alternatively an approach of CBR modelling designed as a closed loop process. The main characteristic of this new approach is to consider and evaluate user feedback and "satisfying" data and applying this data and user information in a closed loop cycle for dynamically optimising the relationship and assignment functions within the problem/solution pairs.
The procedure using statistically evaluated feedback data (fitness function) assures implicitly the consistency of the case base (Fig.3) and furthermore optimises the adaptation of known "old" solutions to current problems resp. the generation of new solutions. This presented CBR methodology is discussed and illustrated as an enhancement of a decision support strategy and decision making process already proposed in several papers [1, 2, 3, 10, 11] at the Euromedia conferences 2000-2002:

The main approach of this CBR modelling (in the following called DSU - Decision Support Unit) consists in a dynamically relationship between subject trees (SCT - the case base, consisting of solutions) and task trees (TCT - the problem base - consisting of problems resp. tasks), whereby the TCT is structured in the same way as the SCT. As illustrated in (Fig.4) the correspondence between SCT and TCT is realised as a structured relationship function SRF, which implies similarity terms and coefficients, feedback terms, calculated out of the feedback data given by the user and further coefficients, calculated out of the statistical analysis of the system history data like the accumulated satisfying factor (ASF, TC, SF) [2, 3]. The dynamically growing and expansion of the TCT is referenced to the SCT by the varying relationship function SRF - the TCT is organically mapped with the SCT. Each object in the TCT is formalized as a Structured Task Description Vector - STDV. A permanently running module, the relationship module, calculates the relationship function out of the data given within questionnaires, filled in by the user, similarity estimations using the node representatives and further statistical history data calculations.
A further main characteristically significance of the DSU is the use of task related multi-loop feedback procedures. The multi loop feedback structure ensures stable generation of dynamical knowledge pools embedded in coupled subject trees - and this in relation to given task-profiles of the users. The correspondences between subject (object) trees and user specified task (- profile) trees are adapted (and varying) according to multi loop feedback evaluations. This guarantees a dynamically growing and implicitly supervised corresponding knowledge trees (the case bases - for problems/tasks and solutions):

Fig.5: External feedback: the user evaluates the usefulness of recommended objects with respect

Internal feedback loops based on fitness functions evaluate and determine which parts of the user specified task/problem description are fitting well (and in which degree) to the objects recommended by the system. That means the user is assisted and supported in optimising his task resp problem description. This fitness function (loop in Fig. 4, Fig. 5) also gives implicitly advises to the user to avoid generation of inconsistencies and indicates probable contradictions. A further internal loop considers statistical history data to facilitate modifications of the task description by the user.

Within the external feedback loop the user evaluates recommended objects with respect to and in relation to the user specified task. The user estimates the usefulness of the objects by filling in a questionnaire and gives feedback evaluations. This evaluation and feedback data are employed implicitly for optimising and dynamically adapting the relationship and assignment function SRF between the TCT (Task/Problem Classification Tree) and the SCT (Subject/Solution Classification Tree). This yields in optimised correspondences between the task-profiles and the knowledge objects stored and handled in the SCT. In this sense the DSU provides facilities for a knowledge (information) retrieval process which is instantiated and supervised by a decision making process - supporting and assisting the user resp. researcher in determining and classifying his information requirements related to his given task/problem description. The main target of the DSU is to recommend for the given task a set (list) of approaches, methods, documents, software libraries and other relevant knowledge objects that satisfy the given criteria determined by the user - and furthermore give recommendations for modifications of the task description itself.

The retrieval procedure is undertaken stepwise in an iterative way, so that the user is conducted in optimising his task description and accordingly optimising his information retrieval. The system extracts out of the user"s (researcher"s) predefined task characteristically knowledge patterns (- profiles) and accordingly recommends relevant information objects. These objects are automatically rated with respect to their task/problem related relevancy. According to the illustration given in Fig. 4 the procedure is as follows:

1 - Define/Edit Task: The DSU provides a multi step request cycle to the user in which the information request is undertaken stepwise in an iterative way. The user determines in a 10-step cycle his task and provides the system with data concerning: task name, subject, keywords, narrative task description, aim description, related authors, methods, time constraints, exclude subjects, ex-clude authors etc.

2 - STDV Extraction: The system extracts out of this data a knowledge pattern and generates a user specific task profile, called the Structured Task Description Vector (STDV, see previous paragraph). This profile is stored in the automated generated Task (problem) Classification Tree (TCT) the problem case base.

3 - Check Sim-STDV: The user specified task is checked within an internal loop with respect to already stored and processed history data. If a similar task exists (within a predefined range or threshold) the user accesses the relevant objects and modifies accordingly his task description (internal loop).

4/5 - TCT-SCT-Correspondence and assignment function: The STDV given by the user is checked concerning its similarity and fuzzy matched with the task description tree (TCT). The given structured relationship and assignment function (SRF) establishes a weighted reference to the subject tree.

6 - Recommend Objects: According to the weight of the relationship function a set of knowledge objects is recommended to the user. The objects are rated with a relevancy factor

7 - Fitness Function: Within a further internal loop an implicit fitness function (relevancy checking) evaluates and determine which parts (components) of the user specified task/problem description (STDV) are fitting well (and in which degree) to the recommended objects. Accordingly the user modifies his task description.

8 - Select Objects: The user selects or discards recommended objects with respect to his given task description.

9 - Storing: The recommended objects are stored, referenced and assigned to the given STDV.

10 - Evaluation-Feedback: The user determines and evaluates the usefulness of recommended objects / solutions with respect to her given task description.

11 - Optimise SRF: As an external feedback loop the feedback data are exploited to optimise and adapt the relationship function SRF (as correspondence function) between task and subject trees. The main scientific approach here is that a dependency is established between the object ratings and the predefined task.

The main application target of Web Services is to establish and enhance the interoperability between different information providing and information demanding entities. An essential attribute of Web Services should be platform and language independence, which guarantee the useful integration in heterogeneous environments. In this context several languages define standards for service discovery, description and messaging protocols:

This given web service standards are useful and applicable for generating single and static web services. In case of composing dynamically existing web services the use of this standards is not sufficient. Accordingly further enhancing languages have been proposed and developed, which extends the current web technologies with the focus of the semantic web. Due to the lack of handling dynamically processes in web especially in semantic web technology the OWL - the web ontology language has been specified in 2001 [14, 15, 16] by the W3C et al. Therefore in combination with DAML+OIL (DARPA Agent Markup Language plus Ontology Inference Language) & DAML-S (DAML service language) the standards and languages are presented for handling and managing web services also under aspects of dynamical composing. In [17] a concept and some tools are described for the semi-automatic composition of web services. DAML-S decomposes the semantic description of a web service into three components: the service profile (input / output types), process model (how the service works) and grounding (details of how an agent can access the service). The service composition prototype of [17] has two basic components: a composer and an inference engine, whereby the composer is the user interface that handles the communication between the human operator and the engine. The inference engine, which is realised in [17] as an OWL reasoner built in Prolog, stores the information about known services in its Knowledge Base. At this stage the present authors propose a further new approach and concept: the application of hybrid case based reasoning as a fusion of the Knowledge Base and the inferencing process. The web service Knowledge Base is designed as a typical case (task and/or problem) base, reflecting the abilities and features of stored web services. In conjunction with a simple structured rule base for supervising the matching process between existing and new designed web services the more complicated structure of separate inference engine and Knowledge Base is eliminated. A further advantage is the direct access and the comparison of existing cases (web services) as the assumption for building new (similar) web services. In this sense the composing process is decision supported by using the similarity phenomena, explained in the previous sections.

Meanwhile a European consortium has been formed to continue the ESPRIT/INCO project KNIXMAS in the current 6th Framework. This project will be based on the scientific results the KNIXMAS project and furthermore enhance the former project achievements with modern and contemporary techniques and methods, especially the modification of the CBR approach and the implementation and introduction of Web services using DAML + OIL, OWL, OOA and DAML S. Within this new project the focus will be set to establish a framework for intelligent Web services created by Web Service composers accessing a decision supported organizational memory as the centred module in the framework.

[1] Nern H,-J. et al; "Final Report - Knowledge Shared XPS-Based Research Network Using Multi-Agent Systems", European project: ESPRIT HPCN : 977113 KNIXMAS; http://www.cordis.lu/ esprit/src/977113.htm
[2] Grancharova A., Nern H.-J., et al; "Decision Support Unit as Part of a Global Knowledge-Shared Research Network", Euromedia 2000, May 8-10, 2000, Antwerp, Belgium
[3] Nern H.-J., Grancharova A., et al; "Decision Strategies for Rating Objects in Knowledge-Shared Research Networks", Proc. of 4th World Multi-Conference on Circuits, Systems, Communications & Computers (CSCC 2000), Athen, Greece, July 10-15, 2000,
[4] Atanasova T., Nern H.-J., et al; "Distributed Heterogeneous Knowledge Data Base for Control System Design: Multi-Agent Development and Support", IEEE-ISIC 2000, Int. Symposium on Intelligent Control, Rio, Patras, Greece, 17-19 July, 2000
[5] Bergmann R., Breen S., Goeker M.; "Developing Industrial Case Based Reasoning Applications", the INCREA Methodology, Springer 1999
[6] Inakoshi H., Okamoto S., Ohta Y. "CBR Wrapper: New Framework for Product Configuration", Proc. 1. Conference on Professional Knowledge Management, Baden-Baden, March 2001, Shaker, Aachen, 2001
[7] Sabin, D. Weigel, R. "Product configuration frameworks - a survey", IEEE Intelligents Systems13(4):39-88, 1992
[8] Bergmann R., Richter M.M., Schmitt S.; "Utility-Oriented Matching: A New Research Direction for Case-Based Reasoning", Proc. 1. Conference on Professional Knowledge Management, Baden-Baden, March 2001, Shaker, Aachen, 2001
[9] Althoff K.D., Richter M.; "Similarity and Utility in Non-Numerical Domains", in Mathematische Methoden der Wirtschaftswissenschaften, Physika-Verlag, 1999
[10] Nern H.-J., Rusc T.; "Filtering and Classification of Meta-Data-Information Objects for Knowledge Sharing within an AI-Based Research Network" , Euromedia 2001, 18-20th April 2001, Valencia, Spain
[11] Atanasova T., Nern H.-J., Pautzke, F.; "Multi-Agent Approach for Task Related Decision Supported Information Retrieval", Euromedia 2002, April 2002, Modena, Italy
[12] Christensen E., et al "Web Services Description Language (WSDL)", http:// www.w3.org/TR/2001/NOTE-WSDL-20010315, 1.1, 2000
[13] "The UDDI Technical White Paper", http://www.uddi.org/ , 2000
[14] "W3C.SOAP 1.2", http://www.w3.org/TR/2002/CR-soap12-part0-20021219/
[15] T.Berners-Lee, J.Hendler, O.Lassila ,"The Semantic Web", Scientific American, 284(5), p34-43, May 2001
[16] DAML Services Coalition " DAML: Web Service Description for the Semantic Web", First International Semantic Web Conference (ISWC), June 2002
[17] Evren Sirin, James Hendler, Bijan Parsia, "Semi-automatic Composition of Web Services using Semantic Descriptions", "Web Services: Modeling, Architecture and Infrastructure" workshop in conjunction with ICEIS2003, 2002