Manfred Meyer

ARC-TEC : acquisition, representation and compilation of technical knowledge

A global description of an expert system shell for the domain of mechanical engineering is presented. The ARC-TEC project constitutes an AI approach to realize the CIM idea. Along with conceptual solutions, it provides a continuous sequence of software tools for the acquisition, representation and compilation of technical knowledge. The shell combines the KADS knowledge-acquisition methodology, the KL-ONE representation theory and the WAM compilation technology. For its evaluation a prototypical expert system for production planning is developed. A central part of the system is a knowledge base formalizing the relevant aspects of common sense in mechanical engineering. Thus, ARC-TEC is less general than the CYC project but broader than specific expert systems for planning or diagnosis.

Knowledge-Base Evolution for Product and Production Planning

Knowledge-base evolution techniques are shown to be of critical importance for the successful application of knowledge-based systems in complex domains. By conceptualizing knowledge-base evolution as theory revision, we can take advantage of the basic findings from different research communities. Results from Inductive Logic Programming ILP and Explanation-Based Learning EBL provide a set of techniques that can be used as a foundation for obtaining new knowledge knowledge-base exploration. Techniques from deductive database research might be used for testing the correctness of a knowledge base knowledge base verification. By an interactive application of these exploration and verification techniques, domain experts and other users may similarly improve the effectiveness of the knowledge base knowledge validation. The application of such selected techniques is then discussed with respect to the specific problem of improving production parameters.

A WAM Compilation Scheme

In order to ease the compilation of logic programs, source-to-source transformations are applied to obtain still declarative programs. The horizontal compilation consists of grouping clauses together, performing a flattening process to remove nested compound terms, partially evaluating them, and reordering the constraints followed by variable classification. The constraints are mapped onto a single ≐-primitive.

CoLab: A hybrid knowledge representation and compilation laboratory

Knowledge bases for real-world domains such as mechanical engineering require expressive and efficient representation and processing tools. We pursue a declarative-compilative approach to knowledge engineering.While Horn logic (as implemented in PROLOG) is well-suited for representing relational clauses, other kinds of declarative knowledge call for hybrid extensions: functional dependencies and higher-order knowledge should be modeled directly. Forward (bottom-up) reasoning should be integrated with backward (top-down) reasoning. Constraint propagation should be used wherever possible instead of search-intensive resolution. Taxonomic knowledge should be classified into an intuitive subsumption hierarchy.Our LISP-based tools provide directtranslators of these declarative representations into abstract machines such as an extended Warren Abstract Machine (WAM) and specialized inference engines that are interfaced to each other. More importantly, we provide source-to-sourcetransformers between various knowledge types, both for user convenience and machine efficiency.These formalisms with their translators and transformers have been developed as part of CoLab, acompilationlaboratory for studying what we call, respectively, ‘vertical’ and ‘horizontal’ compilation of knowledge, as well as for exploring the synergeticcolaboration of the knowledge representation formalisms.A case study in the realm of mechanical engineering has been an important driving force behind the development of CoLab. It will be used as the source of examples throughout the paper when discussing the enhanced formalisms, the hybrid representation architecture, and the compilers.

μCAD2NC : a declarative lathe-workplanning model transforming CAD-like geometries into abstract NC programs

μCAD2NC is a knowledge-based system generating workplans for idealized lathe CNC machines. It transforms CAD-like geometries of rotational-symmetric workpieces into abstract NC programs, using declarative term representations for all processing steps. The system has been developed using COLAB, a hybrid-knowledge compilation laboratory which integrates the power of forward and backward reasoning (incI. functional programming), constraint propagation, and taxonomic classification. The focus of this work is on exemplifying techniques of the hybrid, declarative COLAB formalisms for the central subtasks of CAD-to-NC transformations.

FIDO: Finite Domain Consistency Techniques in Logic Programming

In this paper we discuss different implementation for FIDO, a logic programming language with finite domain constraints and consistency techniques. These approaches range from meta-interpretation over horizontal compilation (source-to-source transformation) through vertical compilation down to an extended Warren Abstract Machine. We will stress the horizontal compilation approach which already shows promising results, whereas the deeper integration on a lower implementation layer by extending the unification and control instructions of the WAM will finally give much better performance results.

Using Hierarchical Constraint Satisfaction for Lathe-Tool Selection in a Computer-Integrated Manufacturing Environment

We discuss how to treat the automatic selection of appropriate lathe tools in a computer-aided process planning CAPP application as a constraint satisfaction problem CSP over hierarchically structured finite domains. Conceptually, it is straightforward to formulate lathe-tool selection in terms of a CSP; however, the choice of constraint and domain representations and of the order in which the constraints are applied is nontrivial if a computationally tractable system design is to be achieved. Because the domains appearing in technical applications often can be modeled as a hierarchy, we investigate how constraint satisfaction algorithms can make use of this hierarchical structure. Moreover, many real-life problems are formulated in a way that no optimal solution can be found that satisfies all the given constraints. Therefore, to bring artificial intelligence technology into real-world applications it becomes important to be able to cope with conflicting constraints and relax the given CSP until a suboptimal solution can be found. For these reasons, the constraint system CONTAX has been developed, which incorporates an extended hierarchical arc-consistency algorithm together with discrete constraint relaxation and has been used to implement the lathe-tool selection module of the ARC-TEC planning system.

An alternative to theta-subsumption based on terminological reasoning

Clause subsumption and rule ordering are long-standing research topics in machine learning (ML). Since logical implication can be reduced to rule-subsumption, the general subsumption problem for Horn clauses is undecidable [Plotkin, 1971bl. In this paper we suggest an alternative knowledge-representation formalism for ML that is based on a terminologicallogic. It provides a decidable rule-ordering which is at least as powerful as 0-subsumption.

Finite Domain Consistency Techniques: Their Combination and Application in Computer-Aided Process Planning

In this paper we present the weak looking-ahead strategy (WLA), a consistency technique on finite domains combining the computational efficiency of forward-checking with the pruning power of looking-ahead. We show that by integrating weak looking-ahead into PROLOGs SLD resolution we obtain a sound and complete inference rule, whereas standard looking-ahead itself has been shown to be incomplete. We outline how we use the weak looking-ahead technique for lathe tool selection in a CIM environment.

Weak looking-ahead and its application in computer-integrated process planning

Constraint logic programming has been shown to be a very useful tool for knowledge representation and problem-solving in different areas. Finite Domain extensions of PROLOG together with efficient consistency techniques such as forward-checking and looking-ahead make it possible to solve many discrete combinatorial problems within a short development time. In this paper we present the weak looking-ahead strategy (WLA), a new consistency technique on finite domains combining the computational efficiency of forward-checking with the pruning power of looking-ahead. Moreover, incorporating weak looking-ahead into PROLOGs SLD resolution gives a sound and complete inference rule whereas standard looking-ahead itself has been shown to be incomplete. Finally, we will show how to use weak looking-ahead in a real-world application to obtain an early search-space pruning while avoiding the control overhead involved by standard looking-ahead.