The broad concept of "the invisible hand" informs much of modern economic theory. This concept can be adapted for distributed control systems by having each separate agent in the system run a reinforcement algorithm, thereby emulating individual humans in an economy. The idea is to design the reward functions of each of those agents so that a Nash equilibrium of the entire system optimizes the behavior of the entire distributed control system. It turns out these distributed control techniques can also be used for distributed optimization. In that context, the cross-entropy method, genetic algorithms, and more generally evolution of distributions algorithms are just special cases. We can then exploit a formal correspondence between distributed optimization and machine learning to improve these distributed optimization algorithms, resulting in (better than) state of the art performance.
Back to Workshop III: Naturalistic Approaches to Artificial Intelligence
Copyright. All Rights Reserved.
