Learning in Multiagent Control of Smart Matter

Embedding microscopic sensors, computers and actuators into materials allows physical systems to actively monitor and respond to their environments. This leads to the possibility of creating smart matter, i.e., materials whose properties can be changed under program control to suit varying constraints. A key difficulty in realizing the potential of smart matter is developing the appropriate con¯ ol programs. One approach to this problem is a multiagent control system, which can provide robust and rapid response to environmental changes. To improve the performance of such systems, we describe how the agents’ organization can adapted through simple learning mechanisms. As a specific example, we consider maintaining a physical system near an unstable configuration, a particularly challenging application for smart matter. This shows how the system’s organization can adapt to the local physical structure to improve performance.