Foundational Qualitative Domain Theories


One hallmark of common sense reasoning is the ability to draw useful inferences even with little information. Research on mental models suggests that qualitative representations are an essential component of common sense reasoning about the physical world. This common sense qualitative knowledge also provides the framework for efficient application of more quantitative and/or domain-specific knowledge. Qualitative reasoning orchestrates the use of more detailed knowledge by detecting what important questions remain unresolved, and by identifying the relevant phenomena which must be investigated to resolve them. Because of this, qualitative reasoning is seeing widespread application in engineering tasks.

We believe qualitative physics can provide similar payoffs for the tasks driving the HPKB initiative. In JFACC campaign planning, for example, an important capability is estimating the indirect effects of actions. Traditional planning representations encode the direct effects of an action on the environment. But they do not support reasoning about the indirect physical effects of actions, especially the changes wrought by physical processes. The ability to support reasoning about the effects of actions on complex engineered systems, even with little detailed information, could provide valuable support to effects-based planning.

Qualitative physics can similarly provide payoffs for the Genoa Intelligence Analysis task. Physical metaphors and simple qualitative models are often combined with other kinds of knowledge to describe the economic aspects of a country. For example, a rising population combined with a poor harvest means either starvation, importing food, or adventurism to gain the necessary resources for survival. This common sense inference rests on treating the population as an entity that requires flows of certain inputs (e.g., food) to remain in existence, knowing what can disrupt those flows (e.g., poor harvest), and knowing what such entities are likely to do when faced with those sorts of conditions (which requires combining this qualitative knowledge with other common sense knowledge of people, economics, and geopolitics). Qualitative representations provide the appropriate level of resolution for situations involving significant uncertainty, because they neither impose unrealistic computational requirements (i.e., accurate numerical data and mathematical models) nor do they generate misleadingly detailed predictions (i.e., a floating point number predicting how many days until a revolution, given the number of inches of rainfall that year). The ability to quickly identify classes of behaviors and their causes, coupled with the ability to characterize the conditions that might lead to one behavior rather than another, could be valuable both in detecting nascent crises and in suggesting courses of action. Returning to our example, a prediction of bad weather over a region, combined with knowledge of its demographics and economic state, can thus lead to the detection of potential political instability and the identification of courses of action that could mitigate the problem (i.e., negotiate trade arrangements to relieve the pressure on the food supply). Tools to craft and trigger indicators for such situations that are based on underlying qualitative models similar to those used by people are likely to be easier to use, and their results will be easier to interpret.

We are creating a library of foundational qualitative domain theories that can provide the common sense substrate for these and similar tasks. The bulk of this library will be created by synthesizing the results of our previous research efforts, which have already yielded powerful models of space, quantity, physical processes, and causality. Our starting points include the following:

[For more details, see our papers here.]

While we are starting from a decade-plus accumulated research base, creating an integrated, coherent library of domain theories still requires work. The formalism used for many of these representations were optimized for particular styles of reasoning such as envisioning. We must recode these representations in the newer formalisms we have developed that better support multi-task reasoning. These representations were also optimized for a particular class of applications, e.g., engineering problem solving. We must do new representation work to fill in the gaps between these theories, to achieve the broad coverage required for common sense reasoning.

To overcome the brittle encodings of these ideas, we will recode them in an augmented axiomatic framework. Specifically, we will use a full predicate calculus language (e.g., KIF) augmented with higher-order representation conventions that provide more natural encodings for particular classes of concepts. More specifically, we will use compositional modeling to express much of the non-spatial content of these domain theories. Compositional modeling organizes domain knowledge into model fragments, which are pieces of knowledge about a phenomenon, principle, law, or equation and modeling knowledge, which expresses relevance information. Given a situation to be represented in the context of some task, a model of that situation is composed by automatic model formulation algorithms that use modeling assumptions to reason about which aspects of the domain theory are relevant to the task at hand.

The methodology of compositional modeling meshes well with the goals of the HPKB program. The abilities to incrementally add detail, to shift perspective and granularity, and to specify control knowledge separately from the domain physics all provide critical support for using foundational domain theories in task-specific knowledge bases.

Model formulation algorithms are still an area of active research, but there are already sufficient useful results. We will use CML (Compositional Modeling Language), a representation language we developed jointly with Xerox PARC, Stanford KSL and UT Austin, in writing our domain theories when possible. CML is a work in progress; the state of the art is evolving too quickly to expect it to remain static. Any extensions we find necessary to CML based on this project will be merged into the specification as appropriate.

Our methodology for creating representations includes rigorous, incremental testing. We will use our existing reasoning systems for this purpose. For example, we will use Gizmo Mark2, a qualitative reasoner designed to provide efficient, incremental qualitative reasoning to test these models. We will also use our diagrammatic reasoning tools in conjunction with Gizmo Mk2 to test the spatial aspects of our representations.

To focus our domain modeling, we will draw on examples of concern to JFACC and GENOA. In addition to the Challenge Problems, we are focusing on theories of weather and climate, civilian logistics distribution systems, and basic properties of vehicles.


Back to HPKB page | Back to QRG Home Page