Commonsense Physics: A Review

Abstract

ion is a two-edged sword. While these abstract state descriptions succinctly capture possible behaviors, they tend not to prescribe exactly which behavior will occur next. By themselves they typically cannot, for we have thrown away just the information required to settle such questions. Thus qualitative simulations tend to be ambiguous. Often such ambiguous answers are good enough: If a household robot cannot imagine any way for the house to burn down as a consequence of its plan to cook supper, it can be satisfied that the plan is at least minimally safe. However, if a house fire is a possibility, more knowledge needs to be invoked. The ability of qualitative physics to represent this ambiguity explicitly is beneficial, since it provides a signal to indicate when more detailed knowledge is required. Exploring the trade-off between abstraction and precision is one of the principle themes of qualitative physics research today. A central goal of qualitative physics is to achieve a degree of systematic coverage and uniformity far in excess of today's knowledge-based systems. In today's expert systems, knowledge is encoded about a particular domain for a particular purpose. Instead of continuing to build such systems, qualitative physics strives to create wide-coverage, multi-purpose domain models. By wide-coverage, we mean that there is some large but precisely characterizable set of systems that can be described by the domain model. It is assumed that every model for a specific system is built by instantiating appropriate elements of the domain vocabulary in appropriate ways. This will reduce the amount of hand-crafting required for new programs, and will hopefully lead to off-the-shelf knowledge bases. By multi-purpose, we mean that a domain model (or a model for a specific situation) can be used for more than one inferential task. Charac­ terizing these styles of reasoning is another goal of qualitative physics. These styles of reasoning include qualitative simulation, interpreting measurements, planning, comparative analysis, and others. Developing domain-independent characterizations of these styles will hopefully lead to generic algorithms that can be used as modules in a variety of larger systems. The literature in qualitative physics has begun to grow rapidly in the last two years, and in order to focus on the new I must slight the old. While this article is self-contained, the reader who wishes to delve more deeply into the basics of the area should see Bobrow ( 1984) and Hobbs & Moore ( 1985), which contain representative samples.