|
|
3. TECHNIQUES AND METHODOLOGIES
3.1. INTRODUCTION
The central concept of model-based reasoning is of course the concept of a model. We therefore start this section by a discussion of model categories and their properties. Selecting a model and a reasoning algorithm for a given purpose are not independent choices, but we treat the latter in a separate subsection.
The term model is generally overused, and AI and the expert system field are no exceptions. In this chapter we link it to an "engineering" notion of a system as an entity that we wish to consider as a unity, whose structure can be decomposed into subsystems and components, and which exhibits certain behavior, i.e., given some input stimuli, it will respond with certain outputs. Based on this notion of a system, a model is a description of a system, using an appropriate modeling language, that makes explicit the structure and behavior of the system for purposes of analysis, prediction, diagnosis, etc.
Some researchers include the function (or purpose) of the system (Chandrasekaran, 1994), in addition to structure and behavior, in their system model. This allows certain types of reasoning that depend on the purpose of the system being studied. For example, in diagnosis, understanding the functional role of a system may help ascertain the impact of a fault and thereby help prioritize and direct the diagnostic search process. In this chapter, we will not include the functional model aspect further.
3.2. MODEL CATEGORIES
Our definition of what a model is, does not prescribe a specific modeling language or representation. Choices of model representation will depend on the purpose of the model, and on available knowledge, resources, tools, and reasoning engines. In defining the space of model categories, we use the distinctions given by Leitch (1992):
- Declarative vs. Procedural. Declarative representations describe relationships between phenomena in the real world, without implying a directionality of the relationship. An example is Ohm's law, V = I * R, which can be used to calculate any of the variables V, I, or R, given values of the other two variables. The lack of directionality makes a declarative representation very general, but can lead to combinatorial complex and inefficient reasoning algorithms. If the available knowledge contains a strong element of directionality, then a procedural representation may be more appropriate. A procedure will efficiently derive the output, given an input. The directionality may however be specific to a certain task or situation, and procedural models therefore tend to be highly specialized. The trend has been to make representation more and more declarative/general, at the expense of applying more computation power (for which the cost is steadily decreasing).
- Quantitative vs. Qualitative. In traditional engineering disciplines, all models are numerical, such as algebraic- or differential equation-based models. Very sophisticated methods for developing and applying such models are available. In AI, researchers have been interested in how models can be used in situations where available knowledge or resources do not enable or justify full quantification, but where qualitative models may suffice (Weld and de Kleer, 1990). In a qualitative model, model variables take on values like "low," "normal," and "high" instead of precise numerical values. Novel representation techniques, including approximate reasoning, fuzzy sets, and order of magnitude, have been developed in this strand of research. In addition to being better tuneable to real availability of domain knowledge, a qualitative model may better reflect a "human" point of view. Humans are good at making rough estimates and predictions of (everyday) physical phenomena based on nonnumerical reasoning.
- Certain vs. Uncertain. The distinction between certain and uncertain knowledge should not be confused with the quantitative vs. qualitative distinction; a mode can well be quantitative and uncertain, or qualitative and certain. However, if the knowledge is uncertain, this fact should somehow be represented in the model. Two main forms of uncertainty have been recognized. The first, probability theory, concerns the situation when exact knowledge is not available, and estimates based on frequency of occurrence are used. The second approach is to represent imprecision directly, for example, as a graded membership of a fuzzy set, a real number between 0 and 1.
- Static vs. Dynamic. Static models only represent steady-state, or equilibrium, behavior of systems. Dynamic models are in addition able to represent model behavior in transient states, as an evolution of state values over time. Until recently, most AI-based representations were based on static models. Such models can be useful, for example, in steady-state diagnosis. However, a dynamic model is essential for proper handling of, for example, control or diagnosis problems when the system is undergoing transient behavior. Dynamic models require the representation of state and storage of matter and energy, and hence delay, that occurs in the physical world. The choice of static or dynamic models fundamentally affects the representation language. In the former, algebraic equations will suffice, whereas in the latter, differential primitives are required.
- Continuous vs. Discontinuous. This final distinction records the fact that in some systems, behavior trajectories can only evolve smoothly through "adjacent" states. This is a characteristic of most systems in nature. However, manmade systems often display discrete state transitions, "jumping" from one state to another state with completely different properties. The latter type of behavior can, for example, be represented by finite-state automata.
The characteristics listed above define a space of possible models for use in model-based reasoning systems. Some of the positions in this space are densely populated, while others have not yet been explored. For example, procedural, quantitative, certain, dynamic, and continuous models represent well-established engineering practise (differential equation-based models), while declarative, qualitative, uncertain, dynamic, and continuous models are intensely investigated in AI. Since the latter category is of particular importance in expert system applications of MBR, we now give a brief overview of the most important developments.
|