Automatic Construction of Accurate Models of Physical Systems

摘要

System identification (SID) is the task of deducing the internal dynamics of a black box system solely from observations of its outputs. This is an essential first step in a variety of engineering problems; most traditional control theoretic methods, for example, require an accurate ordinary differential equation (ODE) model. Accuracy is not the only requirement, however; for efficiency reasons, engineers work hard to construct minimal models; ODEs that ignore unimportant detail and capture only the behavior that is important for the task at hand. This tandem of accuracy and abstraction is a subtle and difficult part of an engineer's training and practice. The computer program PRET, the topic of this report, automates the process described in the previous paragraph by building an artificial intelligence (AI) layer around a set of traditional system identification techniques. This AI layer executes many of the high level parts of the SID process that are normally performed by a human expert, intelligently assessing the task at hand and then reasoning from that information to automatically choose, invoke, and interpret the results of appropriate lower level techniques. These tactics guide PRET quickly and accurately to the minimal ODE that accounts for the important behavior of the system. The ultimate goal of the PRET project is to build a tool that can automatically construct models of high dimensional, nonlinear, black box systems, drawn from any domain that admits ODE models.