Automatic Construction of Accurate Models of Physical Systems
Project Description
PRET automates the process that control theorists call system identification: deducing the internal dynamics of a black-box system solely from observations of its outputs. PRET builds ordinary differential equation (ODE) models of physical systems, linear or nonlinear; it accomplishes this by wrapping a layer of AI techniques around a core of traditional system identification (SID) methods. This AI layer executes many of the high-level parts of the SID procedure that are normally performed by a human expert: it intelligently assesses the situation at each stage of the process and then reasons from the available information to automatically choose, invoke, and interpret the results of appropriate lower-level techniques. During its parameter estimation phase, for example, PRET uses qualitative reasoning techniques to derive good starting values for a nonlinear least-squares solver call, allowing the latter to avoid local extrema in the regression landscape. Like human experts, PRET uses a heterogeneous collection of reasoning modes during the model-building process, and the intelligent orchestration of these modes is critical to its success. This knowledge representation and reasoning task is performed by a special first-order logic system, which selects and coordinates the appropriate reasoning tactics in order to guide the search quickly and accurately to an ODE model that accounts for the important behavior of the system.
PRET is designed to be an engineer’s tool; because of this, it differs sharply from other AI modeling programs in a variety of important ways. First, it explicitly avoids all attempts to “discover” any physics that falls outside its user’s specifications; rather, it works very hard to find a minimal model - one that matches the observations to within a user-prescribed resolution, and no more. Second, it does not adhere to a single, neat paradigm; rather, it calls upon a wide variety of reasoning techniques, ranging from classic AI ideas like constraint propagation to standard engineering tricks like power series, and it works hard to use the right technique at the right time. This mix of methods is the source of PRET’s ability to solve real-world problems in a variety of engineering domains. Third, PRET works directly with the physical world, using sensors and actuators to interact with its target systems — an input/output modeling approach that is both very powerful and extremely difficult because of the nonlinear control theory that is involved.
In terms of performance, PRET attained the functional level of a smart undergraduate. It solves garden-variety textbook SID problems fairly well, struggles occasionally with the harder ones, and has successfully (with some minor hand-holding concerning the relationships between different coordinate systems) modeled one real-world system: a radio-controlled car destined for use in a soccer-playing robot project. SID is an essential first step in the design of a system like this; without an accurate ODE model of the car’s dynamics, it is impossible to build a controller to direct its behavior according to a plan - and ODEs are not part of a Radio Shack spec sheet. From an engineering standpoint, modeling this device is a nontrivial accomplishment; nonlinear SID is considered to be an open problem. The AI issues in this example were also interesting: PRET not only duplicated the model that the project analyst had derived by hand; interacting with the AI tool also helped the human expert refine his explicit mental model of the system.
People
- Prof. Liz Bradley, project leader.
- Matt Easley, who finished his Ph.D. in 2000, developed representation and reasoning facilities for model building and input/output modeling.
- Apollo Hogan, an undergraduate research assistant, wrote PRET's GUI and helped Reinhard with lots of other bits and pieces.
- Brian LaMacchia helped write some of PRET's symbolic algebra facilities.
- Janet Rogers and Abbie O'Gallagher of NIST helped write PRET's nonlinear parameter estimator.
- Reinhard Stolle, who finished his Ph.D. in 1998, designed and implemented PRET's model tester and its declarative reasoning techniques. He also played a key role in most of the main design decisions on the project as a whole.
- Tom Wrensch wrote PRET's qualitative simulator.
Papers
For space reasons, sets of closely related papers have been pruned. See my publications page for a complete list, and please contact me for reprints if what you want isn’t here (or won’t download).
-
General:
- R. Stolle, M. Easley, and E. Bradley, "Reasoning about Models of Nonlinear Systems," in Logical and Computational Aspects of Model-Based Reasoning, L. Magnani et al., eds. Kluwer, 2002.
- E. Bradley, M. Easley, and R. Stolle, "Reasoning about nonlinear system identification," Artificial Intelligence 133:139-188 (2001)
- E. Bradley and R. Stolle, "Automatic Construction of Accurate Models of Physical Systems," Annals of Mathematics and Artificial Intelligence, 17:1-28 (1996).
-
Knowledge representation and reasoning:
- R. Stolle, A. Hogan, and E. Bradley, "Agenda Control for Heterogeneous Reasoners," Journal of Logic and Algebraic Programming 62:41-69 (2005)
- M. Easley and E. Bradley, "Incorporating Engineering Formalisms into Automated Model Builders," in The Computational Discovery of Communicable Knowledge, L. Todorovski and S. Dzeroski, eds. Springer, 2004.
- R. Stolle and E. Bradley, "Communicable Knowledge in Automated System Identification," in The Computational Discovery of Communicable Knowledge, L. Todorovski and S. Dzeroski, eds., Springer 2004.
- M. Easley and E. Bradley, "Meta-domains for automated system identification," Smart Engineering System Design (ANNIE) St. Louis; November 2000. Also available as Department of Computer Science Technical Report CU-CS-904-00
- M. Easley and E. Bradley, "Generalized physical networks for model building," Proceedings International Joint Conference on Artificial Intelligence (IJCAI), Stockholm, August 1999.
- R. Stolle and E. Bradley, "Multimodal Reasoning for Automatic Model Construction", Proceedings Fifteenth National Conference on Artificial Intelligence 1998 (AAAI-98), Madison, Wisconsin, July 1998.
- R. Stolle and E. Bradley, "Opportunistic modeling," Proceedings of the IJCAI (International Joint Conference on Artificial Intelligence) Workshop on Engineering Problems in Qualitative Reasoning, Nagoya Japan; August 1997.
-
Input/output modeling:
- M. Easley and E. Bradley, "Information Granulation in Automated Modeling," in W. Pedrycz, editor, Granular Computing: An Emerging Paradigm, Physica-Verlag, 2001.
- M. Easley and E. Bradley, "Intelligent Sensor Analysis and Actuator Control," IDA-01 (International Symposium on Intelligent Data Analysis), Lisbon; September 2001.
- M. Easley and E. Bradley, "Reasoning about input-output modeling of dynamical systems," International Symposium on Intelligent Data Analysis (IDA), Amsterdam; August 1999.
- E. Bradley and M. Easley, "Reasoning About Sensor Data for Automated System Identification," Intelligent Data Analysis: An International Journal, volume 2, number 2, Elsevier Science (1998).
- Nonlinear parameter estimation:
- E. Bradley, A. O'Gallagher, and J. Rogers, "Global Solutions for Nonlinear Systems using Qualitative Reasoning," Annals of Mathematics and Artificial Intelligence 23:211-228 (1998)
Links
- The qualitative reasoning/qualitative physics home page.
Support
- This material is based upon work supported by the National Science Foundation under grant numbers MIP-9403223 and National Young Investigator Award CCR-9357740, by the Office of Naval Research under grant number N00014-96-1-0720, and by a Packard Fellowship in Science and Engineering. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of these funding agencies.