Physical agents (such as wheeled vehicles, UAVs, hovercraft, etc.) with simple control systems are often sensitive to changes in their physical design and control parameters. As such, it is crucial to evaluate the agent's control systems together with the agent's physical implementation. This can consequently lead to an explosion in the parameter space to be considered. In this paper we investigate the use of swarms of ultra-low complexity agents, and address the issue of finding workable physical agent parameters. We describe a technique for reducing the dimensionality of the search space by performing evaluation tasks that can be used to predict near-optimal parameter values for agents in related multi-agent tasks. We validate our approach on an example task, and demonstrate that this technique can greatly reduce the computational resources required to design a multi-agent system.
@inproceedings{connaughtonetal08aamas,
title={Physical parameter optimization in swarms of ultra-low complexity agents},
author={Ryan Connaughton and Paul W. Schermerhorn and Matthias Scheutz},
year={2008},
booktitle={AAMAS (3)},
publisher={IFAAMAS},
pages={1631--1634}
url={https://hrilab.tufts.edu/publications/connaughtonetal08aamas.pdf}
}