Knowledge Level Description Research Articles

Problem solving methods and knowledge systems: A personal journey to perceptual images as knowledge

AbstractI was among those who proposed problem solving methods (PSMs) in the late 1970s and early 1980s as a knowledge-level description of strategies useful in building knowledge-based systems. This paper summarizes the evolution of my ideas in the last two decades. I start with a review of the original ideas. From an artificial intelligence (AI) point of view, it is not PSMs as such, which are essentially high-level design strategies for computation, that are interesting, but PSMs associated with tasks that have a relation to AI and cognition. They are also interesting with respect to cognitive architecture proposals such as Soar and ACT-R: PSMs are observed regularities in the use of knowledge that an exclusive focus on the architecture level might miss, the latter providing no vocabulary to talk about these regularities. PSMs in the original conception are closely connected to a specific view of knowledge: symbolic expressions represented in a repository and retrieved as needed. I join critics of this view, and maintain with them that most often knowledge is not retrieved from a base as much as constructed as needed. This criticism, however, raises the question of what is in memory that is not knowledge as traditionally conceived in AI, but can support theconstructionof knowledge in predicate–symbolic form. My recent proposal about cognition and multimodality offers a possible answer. In this view, much of memory consists of perceptual and kinesthetic images, which can be recalled during deliberation and from which internal perception can generate linguistic–symbolic knowledge. For example, from a mental image of a configuration of objects, numerous sentences can be constructed describing spatial relations between the objects. My work on diagrammatic reasoning is an implemented example of how this might work. These internal perceptions on imagistic representations are a new kind of PSM.

Neural modeling in cerebral dynamics

For many years our current conception of neural modeling at cortical level has been marked by three convictions: (i) the representational, explanatory, and predictive capabilities of a model are always limited by the mathematical nature of the formal tools used in its formulation. The external observer in fact injects the additional knowledge, apparently resident in the model. (ii) The analogical and logical languages (and consequently, the neurophysiological data which serves as their basis) are not sufficient to describe, model and predict the most genuine aspects of the cortical determined behavior. (iii) We need new conceptual and formal tools capable of representing cooperative processes (not only physical interactions) in cerebral dynamics. In this paper, we use the neuropsychological findings on the residual function after traumatic and surgical lesions in animals and men, to think about the sort of requirements that are necessary to build these formal tools adequate to model neural information processing at cortical level. If we search for inspiration in the field of computation, we arrive to the conjecture that the cerebral dynamics is a dynamics of neurophysiological symbols and, consequently, we need a set of descriptions at an intermediate level (neural assemblies “programming”), in a similar way as we use programming languages at an intermediate level (the symbol level) between the physical machine and the knowledge level descriptions, in the sense of Newell and Marr.

On-line industrial supervision and diagnosis, knowledge level description and experimental results

This paper presents a detailed description of a knowledge-based system for on-line supervision and diagnosis of industrial continuous processes: TURBOLID. The system has been developed for a Spanish beet sugar factory, and it is in use in two plants. The system supports three main tasks; monitoring, operation mode and diagnosis. A detailed knowledge level account of the systems is presented in this paper, using CommonKADS methodology. The knowledge level description presented here will allow the reuse of the TURBOLID problem solving approach to supervision and diagnosis in other continuous plants, even in other domains. The main purpose of this paper is to illustrate how the basic tasks proposed are able to cope with supervision and diagnosis of a complex plant. Particularly significant are the separation of operation mode and diagnosis, and the approach of TURBOLID to diagnosis, that obtains on-line causal explanations for detected problems. TURBOLID is able to differentiate among competing hypotheses looking for historical data, current data and even future data, tracking the plant evolution. Results of the activity of the system during a working month are also presented.

Explanation-based learning: A knowledge level analysis

This paper presents an overview of explanation-based learning (EBL) where the descriptions of EBL methods are realized at the knowledge level. Knowledge level description is a proposal that emphasizes the knowledge content and usage and abstracts away implementation details. Analyzing six different EBL methods at the knowledge level it can be shown what is their shared structure and illuminates their differences in a common framework of description. To describe EBL we use the Components of Expertise (Steels 1991), a knowledge level methodology currently used to describe, analyze, and support the development of expert systems.

SOME METHODOLOGICAL ASPECTS OF MACHINE LEARNING

There is a lack of theory and methodology for the development and proper evaluation of the machine learning (ML) applications built so far. This paper presents some methodological considerations concerning a particular ML perspective expressed in a methodological model called 2D-L1M, a multilevel paradigm-independent methodology that is intended to help during the continuous specification mechanism implicitly present on developing learning programs. Taking into account the striking differences between the knowledge level description of these systems and the symbolic ones, we have identified new layers for their characterization, from the observer point of view and from a pure computational perspective. Our proposal is supported with symbolic structures and arguments from distinguished machine learning strategies and applications (Winston et al., 1983; Mitchell et al, 1983, 1986; Dejong and Mooney, 1986; Carbonell, 1986; Quinlan, 1986; Hammond, 1986; Minton, 1990; Bergadano and Giordana, 1988; Hirsh 1990) and from our own experience in the field (Boticario, 1988; Dent et al., 1992).

Case-Based Reasoning: Foundational Issues, Methodological Variations, and System Approaches

Case-based reasoning is a recent approach to problem solving and learning that has got a lot of attention over the last few years. Originating in the US, the basic idea and underlying theories have spread to other continents, and we are now within a period of highly active research in case-based reasoning in Europe, as well. This paper gives an overview of the foundational issues related to case-based reasoning, describes some of the leading methodological approaches within the field, and exemplifies the current state through pointers to some systems. Initially, a general framework is defined, to which the subsequent descriptions and discussions will refer. The framework is influenced by recent methodologies for knowledge level descriptions of intelligent systems. The methods for case retrieval, reuse, solution testing, and learning are summarized, and their actual realization is discussed in the light of a few example systems that represent different CBR approaches. We also discuss the role of case-based methods as one type of reasoning and learning method within an integrated system architecture.

What you say is what you see ? Interactive generation, manipulation and modification of 3-D shapes based on verbal descriptions

The advent of virtual reality (VR) introduced a paradigm for human-to-human communication in which 3-D shapes can be manipulated in real time in a new kind of computer supported cooperative workspace (CSCW) (Takemura and Kishino 1992). However, mere manipulation — either with 3-D input devices (e.g., the DataGlove™) or with spoken language (Mochizuki and Kishino 1991) — does not do justice to this new paradigm, which could prove to be revolutionary for human-to-human and human-to-machine — communication. This paper discusses the possibility of providing the means for VR-based CSCW participants not only to interactively manipulate, but also to generate and modify 3-D shapes using verbal descriptions, along with simple hand gestures. To this end, the paper also proposes a framework for interactive indexing of knowledge-level descriptions (Newell 1982, Tijerino and Mizoguchi 1993) of human intentions to a symbol-level representation based on deformable superquadrics (Pentland 1986; Horikoshi and Kasahara 1990, Terzopoulos 1991). This framework, at least, breaks ground in integration of natural language with interactive computer graphics.

A stratified approach to specifying, designing, and building knowledge systems

The application of the Soar architecture to building stratified system-level models of knowledge systems is discussed. The stratified approach describes a knowledge system using three levels of computational models: the knowledge-level computational model, the problem-space computational model, and the symbol-level computational model. At the knowledge level, a system is specified in terms of the task it is to perform, the actions it can use to perform the task, and the knowledge it uses to select actions. The knowledge-level description specifies only what the system should do. How the system accomplishes the task computationally is described at the problem-space level as a hierarchy of goals and methods. The system is implemented at the symbol level in Soar using data structures and processes to implement the problem-space design.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

Use of KADS to create a conceptual model of the ONCOCIN task

In recent years, builders of expert systems have become increasingly interested in the Knowledge Acquisition and Design Structuring (KADS) method of knowledge acquisition. Despite the apparent face validity of the KADS approach, reports of the use of KADS to develop large application systems are virtually non-existent. We used KADS to model the task of cancer-chemotherapy administration performed by ONCOCIN, a large medical expert system developed at Stanford University in the 1980s. Based on a knowledge-level description of ONCOCIN and on a review of the features of KADS, we developed K-ONCOCIN, a model of the ONCOCIN cancer-chemotherapy task that can be understood in terms of the four layers required by the KADS approach. In this paper, we offer a detailed description of the elements and layers of the K-ONCOCIN conceptual model. We also describe briefly a design model that implements K-ONCOCIN as a functional expert system. The K-ONCOCIN model provides a data point for understanding the strengths and limitations of the KADS method when applied to a complex domain task.

Modelling Expertise for KBS Development

The benefits of a qualitative modelling approach to building knowledge-based systems include better project control, communication between participants, more accurate validation and easier maintenance. Knowledge acquisition from experts will be more successful if a conceptual model is developed from a knowledge level description, to act as a publicly examinable basis for system design. Analytical tools are available to structure both content and process knowledge on paper, before any commitment is made to implementation. The role of abstraction in developing models of generic tasks as a set of templates for conceptual model development is reviewed, with particular regard to knowledge acquisition tools. Modelling problem-solving methods, especially for constructive tasks such as planning, is more difficult than modelling content knowledge. Epistemological support for conceptual abstraction, particularly at middle levels, remains inadequate.

Modelling Expertise for KBS Development

The benefits of a qualitative modelling approach to building knowledge-based systems include better project control, communication between participants, more accurate validation and easier maintenance. Knowledge acquisition from experts will be more successful if a conceptual model is developed from a knowledge level description, to act as a publicly examinable basis for system design. Analytical tools are available to structure both content and process knowledge on paper, before any commitment is made to implementation. The role of abstraction in developing models of generic tasks as a set of templates for conceptual model development is reviewed, with particular regard to knowledge acquisition tools. Modelling problem-solving methods, especially for constructive tasks such as planning, is more difficult than modelling content knowledge. Epistemological support for conceptual abstraction, particularly at middle levels, remains inadequate.

Model-based knowledge acquisition for heuristic classification systems

We present a methodology and an implemented system, EAR*, for supporting knowledge acquisition. Its goal is to build an implementation-independent knowledge-level description of the target knowledge base, guided by our current understanding of the generic structure of heuristic classification systems.