Iterative Adaptive Control: Windsurfing with Confidence

Abstract

Learning, apparetly, is the result of making mistakes. A good example is learning to windsurf. The approach I took is simple. Fall Often. This is the approach I seem to use regularly when learning anything. It helped a great deal that I did this in Tahiti where both the air and water were 80 degrees and there was little wind. I was in Tahiti following my first visit with Brian Anderson in Canberra in 1984. A few years later when I was describing my learning experience at windsurfing to Brian, who by the way is an excellent windsurfer, we realized that how a human learns windsurfing might be a procedure that could be used for adaptive control. We speculated about human learning of physical activities as fallows. The human first learns to control over a limited bandwidth and learning pushes out the bandwidth over which an accurate knowledge of the”plant” is known. The human first implements a low-gain controller and learning allows the loop to be tightened. The human tends to “over-react” when the environment changes, often resorting to high- gain or no-gain. Then falls. The few instants just before falling may provide some new information about the next higher increments in the bandwidth of the system. So of course mistakes are fundamental to learning. (This is why stabilizing feedback is negative; the popular concept of positive feedback may be useful, in fact, when it is actually negative!) Our speculations on this “windsurfing” approach to adaptive control was first expressed in [2] and subsequently filled out in a series of papers [24,25].