Swarm Engineering

Citation

Kazadi, Sanza Nkashama Tsilobo (2000) Swarm Engineering. Dissertation (Ph.D.), California Institute of Technology. doi:10.7907/YR44-KF67. https://resolver.caltech.edu/CaltechTHESIS:10062010-090510798

Abstract

Swarm engineering is the natural evolution of the use of swarm-based techniques in the accomplishment of high level tasks using a number of simple robots. In this approach, one seeks not to generate a class of behaviors designed to accomplish a given global goal, as is the approach typically found in mainstream robotics. Once the class of behaviors has been understood and decided upon, specific behaviors designed to accomplish this goal may be generated that will complete the desired task without any concern about whether or not the final goal will actually be completed. As long as the generated behaviors satisfy a set of conditions generated in the initial investigation, the desired goal will be accomplished. This approach is investigated in terms of three specific practical problems. First we apply swarm engineering to plume tracking, utilizing both real and simulated experiments. We initially decide whether or not this is a problem that swarm engineering should be applied to at all, A careful investigation of two plume-tracking algorithms yields the weakness of single-robot plume tracking — the tracking of low density plumes. Application of swarm engineering to this sub-problem yields performance that is capable of tracking plumes that are significantly more diffuse than those which can be tracked by single plume trackers. The second problem is position-independent clustering, in which a cluster of objects is constructed from many initial clusters, without predetermination of the final cluster's location. Although we have completed a robotic instantiation of this type of system, the bulk of our work is theoretical and verification with simulation. A minimal condition is derived which guarantees this final global outcome, in both an unrealistic case and a case that is more realistic for real robots. Methods of generating efficiency improvements are derived. The application of this formalism to real robots is discussed. Finally, conditions are derived and demonstrated leading to stable arrangements of multiple clusters, a precursor of distributed construction. The third and last problem is the traveling salesman problem (TSP). Marco Dorigo (see, for instance, Dorigo 1996) generated much interest in the use of "ants" on the problem. We theoretically demonstrate that Dorigo's ants satisfy a swarm engineering condition generated for this problem. However, the result of Dorigo's work is also demonstrably fragile under random fluctuations. A second more robust ant-based method is generated and tested on a small number of standard test problems taken from TSPLIB, a source of many well-known TSP instantiations. In all cases, the new algorithm is capable of reliably finding the minimum distance path. This thesis makes a number of contributions to the literature. First, we study two systems that take advantage of the embodiment and interaction dynamics of the robot to accomplish the single robot goal. We then extend this work to answer two questions: 1. Is it possible to apply swarm engineering to plume tracking? 2. How can one apply swarm engineering to plume tracking? We demonstrate that is indeed possible to extend one of the two previous plume tracking systems to a swarm. We demonstrate that the swarm is capable of successfully tracking plumes that are significantly more diffuse than those capable of being tracked by individual agents. Next, we derive a formal condition which must be satisfied in order for a system made up of many clusters and robots to become a system of one cluster and many robots, if the robots do not have global information. We extend this work by deriving ways of generating more efficient paradigms, and demonstrate how to use the theory in the design of individual robots. Finally, we modify our theory to include systems generating multiple clusters, and demonstrate how to generate clusters of predefined relative sizes. Our third contribution is a formalism for ant-based TSP algorithms. This allows us to understand why Dorigo's algorithms are relatively unstable. We create a completely new system of traveling agents that also solve TSP by having the ability to "die along the way" and to self replicate in a specific fashion. Perhaps the most important contribution here is an alternative approach to the design of swarm systems. This allows us to explore the three problems of interest here by first determining a condition that allows the completion of the task, and then allows the generation of behaviors satisfying the minimal condition. This view to system design has the advantage of allowing the design of systems that will generate the desired behavior, concommittantly reducing the danger of generating unwanted emergent behavior.

Thesis Files