System identification takes a space of possible models and a stream of obse
rvational data of a physical system, and attempts to identify the element o
f the model space that best describes the observed system. In traditional a
pproaches, the model space is specified by a parameterized differential equ
ation, and identification selects numerical parameter values so that simula
tion of the model best matches the observations. We present SQUID, a method
for system identification in which the space of potential models is define
d by a semi-quantitative differential equation (SQDE): qualitative and mono
tonic function constraints as well as numerical intervals and functional en
velopes bound the set of possible models. The simulator SQSIM predicts semi
-quantitative behavior descriptions from the SQDE. Identification takes pla
ce by describing the observation stream in similar semi-quantitative terms
and intersecting the two descriptions to derive narrower bounds on the mode
l space. Refinement is done by refuting impossible or implausible subsets o
f the model space. SQUID therefore has strengths, particularly robustness a
nd expressive power for incomplete knowledge, that complement the propertie
s of traditional system identification methods. We also present detailed ex
amples, evaluation, and analysis of SQUID. (C) 2000 Elsevier Science B.V. A
ll rights reserved.