This dissertation is given in two parts followed
by concluding remarks. The first three chapters
describe the generalization of optimal foraging
theory for the design of solitary task-processing
agents. The following two chapters address the
coordinated action of distributed independent
agents to achieve a desirable global result. The
short concluding part summarizes contributions and
future research directions.
Optimal foraging theory (OFT) uses
ecological models of energy intake to predict
behaviors favored by natural selection. Using
models of the long-term rate of energetic gain of a
solitary forager encountering a variety of food
opportunities at a regular rate, it predicts
characteristics of optimal solutions that should be
expressed in nature. Several engineered agents can
be modeled similarly. For example, an autonomous
air vehicle (AAV) that flies over a region
encounters targets randomly just as an animal will
encounter food as it travels. OFT describes the
preferences that the animal is likely to have due
to natural selection. Thus, OFT applied to mobile
vehicles describes the preferences of successful
vehicle designs.
Although OFT has had success in existing
engineering applications, rate maximization is not
a good fit for many applications that are otherwise
analogous to foraging. Thus, in the first part of
this dissertation, the classical OFT methods are
rediscovered for generic optimization objectives.
It is shown that algorithms that are
computationally equivalent to those inspired by
classical OFT can perform better in realistic
scenarios because they are based on more feasible
optimization objectives. It is then shown how the
design of foraging-like algorithms provides new
insight into behaviors observed and expected in
animals. The generalization of the classical
methods extracts fundamental properties that may
have been overlooked in the biological case.
Consequently, observed behaviors that have been
previously been called irrational are shown to
follow from the extension of the classical
methods.
The second part of the dissertation describes
individual agent behaviors that collectively result
in the achievement of a global optimum when the
distributed agents operate in parallel. In the
first chapter, collections of agents that are each
similar to the agents from the early chapters are
considered. These agents have overlapping
capabilities, and so one agent can share the task
processing burden of another. For example, an AAV
patrolling one area can request the help of other
vehicles patrolling other areas that have a sparser
distribution of targets. We present a method of
volunteering to answer the request of neighboring
agents such that sensitivity to the relative
loading across the network emerges. In particular,
agents that are relatively more loaded answer fewer
task-processing requests and receive more answers
to their own requests. The second chapter describes
a distributed numerical optimization method for
optimization under inseparable constraints.
Inseparable constraints typically require some
direct coordination between distributed solver
agents. However, we show how certain
implementations allow for stigmergy, and so far
less coordination is needed among the agents. For
example, intelligent lighting, which maintains
illumination constraints while minimizing power
usage, is one application where the distributed
algorithm can be applied directly.
![[Ted Pavlic]](images/Ted_head_flowers.jpg "[Ted Pavlic]")
