Self-organizing systems have great potential for adaptability, but as complex systems, they can also be prone to unpredictable behavior, cascading failures, and sensitivity to perturbations. Also, designing systems for adaptability may introduce overhead that reduces their performance. This paper investigates the design tradeoffs between adaptability and performance in the context of a box-pushing task. Using a genetic algorithm to optimize a parametric behavioral model, we are able to test systems optimized under different conditions for performance and reliability. It was found that a system optimized in the face of random initial conditions and internal perturbations was more robust than a system optimized without these perturbations, showing a higher overall fitness over diverse trials, but it could not take advantage of particular initial conditions that would have allowed it to achieve a one-off high fitness in a repeatable environment. A system optimized with predesigned initial conditions was found to achieve a very high fitness, but when it was retested with random initial conditions, its performance plummeted, indicating that it was fit to the ideal initial conditions but not robust.

This content is only available via PDF.
You do not currently have access to this content.