1.1.   THE LEGACY OF CONTROL TECHNIQUES

At the time of conception of the DARPA Software Enabled Control (SEC)
program, control research was proceeding down a path determined by old
views of the computational and systems context. The assumptions were as
follows: highly constrained sensing and actuation, limited processing and
communication resources, computational intractability of large or even moderate
state spaces, poorly characterized and unpredictable switching effects,
and target systems that operated independently and without interaction with
other systems.

Control theory and engineering have a remarkably successful history of
enabling automation, and information-centric control is by now pervasive.
Yet today’s controllers are conservative: Being products of overdesign, they
often yield underperformance. Their designs are statically optimized for
nominal performance, around simplified time-invariant models of system
dynamics and a well-defined operational environment. They also fail in
unexpected circumstances: control vulnerabilities that arise in extreme environments
are frequently ignored. System modification (reconfiguration, dam-
age, failure) may demand large changes in the controller, perhaps online
during operation.

Research in adaptive control has sought to accommodate change through
the use of online feedback to the governing parameters; robust control
research introduced periodic recalculation of the controller, using model-
based prediction. Relative to simple PID and set-point control, these model-
centric strategies yield improvement. However, there are several problems.
Disruptive events occur unexpectedly, not periodically, and the changes
required may be dramatic. Principles and support are lacking for reactive but
systematic online reconfiguration of models and software. A popular modern
strategy is to enlarge models for ‘‘full-envelope design,’’ attempting to accommodate
an ever-larger span of configurations and environmental conditions
in a single control law. This results in monolithic, flattened models with
immense state spaces.