Family Discovery
Stephen M. Omohundro · 1995
Many learning situations involve multiple sets of training examples drawn from different but related underlying models. "Family dis-covery" is the task of discovering a parameterized family of models from this kind of training set. The task naturally arises in density estimation, classification, regression, manifold learning, reinforcement learning, clustering, HMM learning, and other settings. We describe three family discovery algorithms which are based on techniques for manifold learning. We compare these algorithms on a classification task against two alternative approaches and find significant performance improvement. 1