17.04.2015 Views

DARPA ULTRALOG Final Report - Industrial and Manufacturing ...

DARPA ULTRALOG Final Report - Industrial and Manufacturing ...

DARPA ULTRALOG Final Report - Industrial and Manufacturing ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

• Classification of the signal: System identification in nonlinear chaotic systems means<br />

establishing a set of invariants for each system of interest <strong>and</strong> then comparing<br />

observations to that library of invariants. The invariants are properties of attractor <strong>and</strong> are<br />

independent of any particular trajectory of the attractor. Invariants can be divided into<br />

two classes: fractal dimensions (Farmer et. al. 1983) <strong>and</strong> Lyapunov exponents (Sano <strong>and</strong><br />

Sawada 1985). Fractal dimensions characterize geometrical complexity of dynamics i.e.<br />

how the sample of points along a system orbit are distributed spatially. Lyapunov<br />

exponents on the other h<strong>and</strong> describe the dynamical complexity i.e. “stretching <strong>and</strong><br />

folding” in the dynamical process.<br />

• Making models <strong>and</strong> Prediction: This step involves determination of the parameters of<br />

the assumed model of the dynamics:<br />

y(<br />

n)<br />

→ y(<br />

n + 1)<br />

y(<br />

n + 1) = F(<br />

y(<br />

n),<br />

a , a<br />

,..... a<br />

1 2 p<br />

which is consistent with invariant classifiers (Lyapunov exponents, dimensions). The functional<br />

form F (⋅) often used, includes polynomials, radial basis functions etc. Local False Nearest<br />

Neighbor (Abarbanel <strong>and</strong> Kennel 1993) test is used to determine how many dimensions are<br />

locally required to describe the dynamics generating the time series, without knowing the<br />

equations of motion <strong>and</strong> hence gives the dimension for the assumed model. The methods for<br />

building nonlinear models can be classified as Global <strong>and</strong> Local (Farmer <strong>and</strong> Sidorowich 1987;<br />

Casdalgi 1989). By definition Local methods vary from point to point in the phase space while<br />

Global Models are constructed once <strong>and</strong> for all in the whole phase space. Models based on<br />

Machine Learning techniques such as radial basis functions or Neural Networks (Powell 1987)<br />

<strong>and</strong> Support Vector Machines (Mukherjee et al. 1997) carry features of both. They are usually<br />

used as global functional forms, but they clearly demonstrate localized behavior too.<br />

The techniques from nonlinear time series analysis are well suited for modeling the<br />

nonlinearities in the supply chains. For an application of nonlinear time series analysis in supply<br />

chains, the reader is referred to Lee et al., 2002. Using it one can deduce that the time series is<br />

deterministic, so that it should be possible in principle to build predictive models. The invariants<br />

can be used to effectively characterize the complex behavior. For e.g., the largest Lyapunov<br />

exponent gives an indication of how far into the future, reliable predictions can be made while the<br />

fractal dimensions gives an indication of how complex a model should be chosen to represent the<br />

data. These models then provide the basis for systematically developing the control strategies. It<br />

should be noted the functional forms used for modeling in the step (4) above, are continuous in<br />

their argument. This approach builds models viewing a dynamical system as obeying laws of<br />

physics. From another perspective a dynamical system can be considered as processing<br />

information. So an alternative class of discrete “computational” models inspired from the theory<br />

of automata <strong>and</strong> formal languages can also be used for modeling the dynamics (Marcus 1996).<br />

“Computational Mechanics”, considers this viewpoint <strong>and</strong> describes the system behavior in terms<br />

of its intrinsic computational architecture i.e. how it stores <strong>and</strong> processes information.<br />

6.2 Computational Mechanics<br />

Computational mechanics is a method for inferring the causal structure of stochastic processes<br />

from empirical data or arbitrary probabilistic representations. It combines ideas <strong>and</strong> techniques<br />

from nonlinear dynamics, information theory <strong>and</strong> automata theory, <strong>and</strong> is, as it were, an “inverse”<br />

to statistical mechanics. Instead of starting with a microscopic description of particles <strong>and</strong> their<br />

interactions, <strong>and</strong> deriving macroscopic phenomena, it starts with observed macroscopic data, <strong>and</strong><br />

infers the simplest causal structure: the “ε -machine” capable of generating the observations. The<br />

ε -machine in turn describes the system's intrinsic computation, i.e., how it stores <strong>and</strong> processes<br />

information. This is developed using the statistical mechanics of orbit ensembles, rather than<br />

)<br />

a j

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!