6.1.   INTRODUCTION

Most control systems built today are resource-limited. This is especially true
for embedded control systems in mobile platforms (whether military or
commercial) due to constraints of size, weight, space, or power. Great effort
is expended in engineering solutions that provide jitter-free periodic execution
while meeting hard real-time deadlines in systems with high CPU
(central processing unit) utilization.

The goal of our research under DARPA’s Software Enabled Control
program has been to develop the software mechanisms that are needed to
allow higher-performance automation and control solutions to be deployed in
embedded, onboard systems. It has been a continuing complaint of the
controls community that sophisticated algorithms that result from research
are usually relegated to the archival literature; what ultimately gets implemented
is a simplified version of the real thing. The exponential growth in
computing capabilities should, in principle, permit implementation of high-
performance compute-intensive algorithms for time-critical applications. But
the middleware and other infrastructural aspects of real-time systems cannot
at present accommodate the novel execution requirements of these algorithms,
and until they can do so the full benefits of Moore’s Law will not be
realized for real-time control.

In this environment, we maintain that a shift from a tightly integrated,
highly coupled design approach to one oriented more toward quality of
service guarantees and tradeoffs is appropriate. For this shift to occur,
real-time platforms must incorporate methods to adapt resource utilization
for computational tasks under unpredictable situations. Models and model-
based algorithms will be central to this capability. A variety of models -
not only representing the physical systems that are being controlled and the
physical environments they are operating in, but also characterizing the
computational performance of processing tasks - must be available for
online interrogation. For example, the active multimodel real-time systems we
envision will perform complex optimizations based on nondeterministic,
incremental processing with variable completion times. The models and
algorithms supporting these optimizations must be resource-aware and adapt
their performance based on the computing time made available to them. The
resource allocation parameters can be treated as additional independent
control variables in the overall optimization problem to achieve the mission
objectives.

Section 6.2 presents some motivating problems for the development and
use of adaptive resource management technology. Sections 6.3 and 6.4 dis-
cuss computational issues and present mechanisms for adaptive allocation of
the on-board computing resources. Section 6.5 introduces the UAV route
adaptation problem that serves to identify and test the computational requirements
for our work on software infrastructure. In Sections 6.6 and 6.7
we apply our computational models to the multiresolution multimodel optimization
of UAV trajectories. Section 6.8 discusses the simulation framework
we have developed to evaluate the performance of our control and computing
models.