Learning to Catch: Applying Nearest Neighbor Algorithms to Dynamic Control Tasks

This paper examines the hypothesis that local weighted variants of k-nearest neighbor algorithms can support dynamic control tasks. We evaluated several k-nearest neighbor (k-NN) algorithms on the simulated learning task of catching a flying ball. Previously, local regression algorithms have been advocated for this class of problems. These algorithms, which are variants of k-NN, base their predictions on a (possibly weighted) regression computed from the k nearest neighbors. While they outperform simpler k-NN algorithms on many tasks, they have trouble on this ball-catching task. We hypothesize that the non-linearities in this task are the cause of this behavior, and that local regression algorithms may need to be modified to work well under similar conditions.