| Model
A simplified version of something complex used, for example,
to analyze and solve problems
Modified Sequential Least Squares
System Identification
A challenge in real time system identification
is the tradeoff between algorithm stability and rapid adaptation
to parameter changes. For reconfigurable flight control, this
is especially important due to long periods of quiescence,
coupled with the possibility of sudden unforeseen failures.
To address these twin challenges, BAI developed the MSLS parameter
identification algorithm. Building on recursive least squares
(RLS), MSLS incorporates additional constraints that take
advantage of a priori information and effectively adjusts
the size of the data-window to account for the information
content in the regressor.
 Reference:
Ward,
D.G, R.L. Barron, M.P. Carley, and T.J. Curtis, "Real-time
parameter identification for self-designing flight control,"
Proc. Nat'l Aerospace and Electronics Conf. (NAECON), Dayton,
OH, May 1994.
Statistical Model Validation
Mathematical modeling and computer simulation
have become indispensable tools in scientific, engineering,
policy, and economic decision-making by government, industry,
and other organizations. Barron Associates’ Stochastic Model
Validation (SMV) technology concerns the use of nonparametric
(distribution-free) statistical techniques for validating multivariate
stochastic simulation models. The validation approach is based
on testing for homogeneity between a population of empirical
data (observed data) and a population generated by one or more
stochastic simulation models. Multivariate techniques are necessary
to reveal whether a simulation model correctly captures dependencies
among multiple response variables. Nonparametric approaches
obviate the need for normality assumptions concerning the distributions
of empirically-observed and model-generated data; in practice,
such distributional assumptions are often not appropriate. Barron
Associates’ SMV approach is novel in that it enables validation
across “multiple treatments,” where empirical and model data
are collected under a variety of “independent test conditions”
that are of operational interest, or which reflect the available
observation data. Integrating non-homogeneous data into one
overall statistical test substantially increases the power of
the test relative to that of testing each treatment separately
and avoids the pitfall of increased probability of significant
findings due to chance alone.
Structure-Learning Neural Networks
BAI's neural network modelling software
suite, GNOSIS, can be used to find the best feedforward model
that transforms observable variables into one or more estimated
dependent variables. GNOSIS postulates a simple model, then
incrementally adds complexity until a "just-sufficient"
complexity level is achieved. In other words, GNOSIS determines
the optimal trade-off between model complexity and model accuracy
when determining the structure of the model. A proper balance
ensures the best performance on unseen data by neither underfitting
nor overfitting the data. BAI has demonstrated the efficacy
of the GNOSIS structure-learning technique in a number of
fields.
Advanced Nonlinear Simulations
BAI works with and develops in-house simulations
of a number of advanced systems ranging from tailless aircraft
to shipboard power distribution systems. As part of these
efforts, BAI is also developing tools to speed-up nonlinear
simulations, to automatically update simulations given experimental
data, and to visualize the results of these simulations.
|
