Machine-generated Explanations

One of the objectives of the How Things Work Project is to develop techniques for generating causal explanations, in natural language, of the simulated behavior of physical devices. A method for generating such explanations has been implemented in the Device Modeling Environment (DME) [3, 4]. DME is a system that helps formulate mathematical simulation models from a library of model fragments using a Compositional Modeling approach, and simulate the modeled systems using qualitative and numerical techniques.

DME generates text and graphics to describe the system being simulated, by composing information from a model library and a trace of the simulation [1, 2]. Although the underlying behavior prediction is accomplished with numerical integration and symbolic reasoning operating on mathematical models, the explanations are abstracted and filtered for human consumption. Causality, salience, and level of detail are determined using domain-independent heuristics. Since these heuristics can't anticipate the exact information needs of the reader, further information is always available. Included with each explanation is a set of follow-up questions. The reader can ask for more information by choosing one of these questions. The path of questions and explanations is generated on demand on the basis of the reader choices.

Example Scenario: Self-explanatory Simulation

We have produced an example scenario in which a moderately complex fluid system is simulated under various conditions to test operator procedures. An engineer has assembled a model of the system using DME, and the results are presented as a large hypertext web of explanations linked by follow-up questions to other explanations. Such a hypertext web can be incorporated into multimedia email conversations.

Consider how such a simulation might be used in an email conversation among engineers. We've mocked up an example email message from a (fictitious) NASA engineer to the the designers of the Reaction Control System used on the Space Shuttles. In the message the engineer refers to a scenario which is a simulation of the RCS and the actions taken by a human operator. The engineer refers to the beginning of the scenario (the pages of the virtual document in which DME summarizes the entire scenario), and also points into particular explanations of specific quantiuties of interest.

How It Works: Virtual Documents

This example demonstrates how the machine-generated, interactive explanation can be delivered in a hypermedia environment such as the World Wide Web. In Web parlance, the product is called a virtual document. In our current implementation, DME acts as an HTTP server that dynamically generates "pages" of explanations in response to "requests" from the http client (e.g., Mosaic). Each request is a query to DME that has been encoded as a virtual document address (a URL). Since DME generates these queries (i.e. the follow-up questions), the user need not worry about the query syntax and can traverse the document using simple user interface gestures such as clicking on highlighted anchors.

The entire web of explanations for this scenario, if saved to disk, would take several hundred megabytes (most of which is text). Thus, this example demonstrates both the potential functionality of machine-generated documentation and the practicality of using a server to deliver virtual documents.

For more information, see the Explanation Generation section of the How Things Work Project Overview, and the references below.

References:

[1]
Gruber, T. R. & Gautier, P. O.(1992). Machine-generated Explanations of Engineering Models: A compositional modeling approach. Proceedings of the 13th International Joint Conference on Artificial Intelligence, Chambery, France, August 1993. pp 1502-1508. Distributed by Morgan Kaufmann Publishers, Inc., San Mateo, CA.
Also available as Technical Report KSL-92-85.

[2]
Gautier, P. O. & Gruber, T. R. (1993). Generating Explanations of Device Behavior Using Compositional Modeling and Causal Ordering. Proceedings of the Eleventh National Conference on Artificial Intelligence. Washington, DC, July 1993. pp 264-270. Published by AAAI Press, Menlo Park, CA.
Also available as Technical Report KSL-93-06.

[3]
Iwasaki, Y. and Low, C.M. Model generation and simulation of device behavior with continuous and discrete changes. Intelligent Systems Engineering, 1(2), 1993.
Formerly KSL-91-69.

[4]
Iwasaki, Y. and Levy, A. Y. Automated model selection for simulation. Knowledge Systems Laboratory, Stanford University, Technical Report KSL-93-11, 1993.

[5]
T. R. Gruber and D. M. Russell. Generative Design Rationale: Beyond the Record and Replay Paradigm. In T. Moran and J. H. Carroll, (Eds.), Design Rationale: Concepts, Techniques, and Use. Lawrence Erlbaum Associates, 1994, in press. Available as Technical Report KSL-92-59.

Tom Gruber <gruber@ksl.stanford.edu>