This paper presents an overview of the INRECA methodology for building and maintaining CBR applications. This methodology supports the collection and reuse of experience on the systematic development of CBR applications. It is based on the experience factory and the software process modeling approach from software engineering. CBR development experience is documented using software process models and stored in different levels of generality in a three-layered experience base. Up to now, experience from 9 industrial projects enacted by all INRECA II partners has been collected.
This paper presents a brief overview of the INRECA-II methodology for building and maintaining CBR applications. It is based on the experience factory and the software process modeling approach from software engineering. CBR development and maintenance experience is documented using software process models and stored in a three-layered experience packet.
For defining attribute types to be used in the case representation, taxonomies occur quite often. The symbolic values at any node of the taxonomy tree are used as attribute values in a case or a query. A taxonomy type represents a relationship between the symbols through their position within the taxonomy-tree which expresses knowledge about the similarity between the symbols. This paper analyzes several situations in which taxonomies are used in different ways and proposes a systematic way of specifying local similarity measures for taxonomy types. The proposed similarity measures have a clear semantics and are easy to compute at runtime.
Complex problem solving can be substantially improved by the reuse of experience from previously solved problems. This requires that case libraries of successful problem solutions are transformed into problem solving knowledge with high utility, i.e. knowledge which causes high savings in search time, high application probability and low matching costs in a respective performance component. Planning can be improved by explanation-based learning (EBL) of abstract plans from detailed, successfully solved planning problems. Abstract plans, expressed in well-established terms of the domain, serve as useful problem decompositions which can drastically reduce the planning complexity. Abstractions which are valid for a class of planning cases rather than for a single case, ensure a successful application in a larger spectrum of new situations. The hierarchical organization of the learned shared abstractions causes low matching costs. The presented S-PABS procedure is an EBL-procedure in which abstraction, learning from multiple examples and hierarchical clustering are combined to automatically construct a hierarchy of shared abstract plans by analyzing concrete planning cases. A specific planning procedure has been designed to solve new planning problems guided by the knowledge learned by S-PABS. By allowing a feedback from this planning procedure to the learning component, the integrated system shows an increase in performance through past problem solving.
Although skeletal plan refinement is used in several planning systems, a procedure for the automatic acquisition of such high-level plans has not yet been developed. The proposed explanation- based knowledge acquisition procedure constructs a skeletal plan automatically from a sophisticated concrete planning case. The classification of that case into a well-described class of problems serves as an instrument for adjusting the applicability of the acquired skeletal plans to that class. The four phases of the proposed procedure are constituted as follows: In the first phase, the execution of the source plan is simulated, and explanations for the effects of the occurred operators are constructed. In the second phase, the generalization of these explanations is performed with respect to a criterion of operationality which specifies the vocabulary for defining abstract operators for the skeletal plan. The third phase, a dependency analysis of the resulting operator effects, unveils the interactions of the concrete plan which are substantial for the specified class. In the forth phase, the concept descriptions for the abstract operators of the skeletal plan are formed by collecting and normalizing the important constraints for each operation that were indicated by the dependencies. With this procedure sophisticated planning solutions from human experts can be generalized into skeletal plans and consequently be reused by a planning system in novel situations.
Abstraction is one of the most promising approaches to improve the performance of problem solvers. Abstraction by dropping sentences of a domain description - as used in most hierarchical planners - is known to be very representation dependent. To overcome these drawbacks, we propose a more general view of abstraction involving the change of representation language. We have developed a new abstraction methodology and a related sound and complete learning algorithm that allows the complete change of representation language of planning cases from concrete to abstract.
Recently, the use of abstraction in case-based reasoning (CBR) is getting more and more popular. The basic idea is to supply a CBR system with cases at many different levels of abstraction. When a new problem must be solved, one (or several) 'appropriate' concrete or abstract case are retrieved from the case base and the solution that the case contains is reused to derive a solution for the current problem, e.g. by filling in the details that a retrieved case at some higher level of abstraction does not contain. A major problem that occurs when using this approach is, that for a given new problem, usually several cases, e.g., from different levels of abstraction could be reused to solve the new problem. Choosing a wrong abstract case can slow down the problem solving process or even prevents the problem from being solved.
Hierachical planning can be improved by explanation-based learning (EBL) of abstract plans from detailed, successfully solved planning problems. Abstract plans, expressed in well-established terms of the domain, serve as useful problem decompositions which can drastically reduce the planning complexity. The learned plan abstraction must be valid for a class of planning cases rather than for a single case, to ensure their successful application in a larger spectrum of new situations. A hierarchical organization of the newly learned knowledge must be archieved to overcome the utility problem in EBL. This paper presents a new formal model of shared plan abstraction and the closely related explanation-based procedure S-PABS. Unlike other apporaches to plan abstraction, our model allows a total different terminology to be introduced at the abstract level. Finally, an unsupervised incremental procedure for constructing a hierachy of shared abstract plans is proposed, as a kind of concept formation over explanations.
As the previous chapters of this book have shown, case-based reasoning is a technology that has been successfully applied to a large range of different tasks. Through all the different CBR projects, both basic research projects as well as industrial development projects, lots of knowledge and experience about how to build a CBR application has been collected. Today, there is already an increasing number of successful companies developing industrial CBR applications. In former days, these companies could develop their early pioneering CBR applications in an ad-hoc manner. The highly-skilled CBR expert of the company was able to manage these projects and to provide the developers with the required expertise.
Der Trend zu einer immer stärkeren Kopplung von Systemen bei gleichzeitiger Dezentralisierung durch Vernetzung hat dazu geführt, daß Computernutzern auf Wunsch enorme Datenmengen zur Verfügung stehen, die sich einer sinnvollen Bearbeitung durch den Nutzer allein völlig entziehen. Unterschiedliche Repräsentationsformalismen für Informationen, Mehrdeutigkeiten, Redundanz sowie eingeschränkte Verfügbarkeit sowohl von Informationen als auch von Rechenleistung machen konventionelle Suchverfahren unanwendbar. Stattdessen werden Suchverfahren und Programme benötigt, die sich intelligent an unterschiedliche Formalismen anpassen, ihre Handlungen ständig evaluieren und fähig sind, ihre Benutzer individuell zu unterstützen. Schlagwörter wie Knowbots, Search-Engines oder Data-Miningsind deshalb zur Zeit in aller Munde. Ein umfassendes Buch, das die hinter diesen und ähnlichen Schlagwörtern verborgenen Ideen und Konzepte präsentiert, existiert jedoch zur Zeit noch nicht. Dies war für uns die Motivation, das Thema "Intelligente Suche im Internet mit Lernenden Systemen" in einem Seminar zu behandeln. Wir haben damit ein Forschungsgebiet aufgegriffen, das sowohl für alle am LSA beteiligten Gruppen von Interesse ist, aber darüber hinaus aktuell von vielen Seiten aufmerksam beobachtet wird. Daher haben wir uns entschlossen, die Ausarbeitungen, die im Rahmen dieses Seminars von den TeilmehmerInnen erstellt wurden, durch den vorliegenden Bericht einer breiteren Öffentlichkeit zugänglich zu machen.