Feature Selection and Hypothesis Selection Models of Induction

Wells, H. (1963). Effects of transfer and problem structure in disjunctive concept formation. Table 2. Mean number of trials required by POSTHOC In POSTHOC, we have focused on how prior knowledge influences learning rates and we have so far ignored other information used by human learners (e.g., perceptual salience of features, Bower & Trabasso, 1968). As a consequence, POSTHOC is not intended to make quantitative predictions on the number of training examples but rather predicts the relative difficulty of learning. Future Directions There are three possible direction in which we plan to extend the hypothesis selection model. First, we would like to be able to use the prior knowledge of the learner to influence the interpretation of ambiguous feature (Medin & Wisniewski, 1990). Second, we would like POSTHOC to be able to use more abstract knowledge. Currently, POSTHOC can represent the information that there there is a specific interaction between two drugs or that there is no drug interaction. In contrast, our subjects also appeared to have more general knowledge that indicates such things as drugs may interact and can use this knowledge to explain the specific interaction seen in the experiment in terms of the general knowledge of drug interactions. Finally, we plan to extend POSTHOC so that when it learns accurate hypotheses that are not consistent with its background knowledge, the background knowledge is revised to accommodate the new findings. Conclusions We have presented experimental evidence that provides support for hypothesis selection models of concept learning. We have extended POSTHOC to include negative influences and shown that with this extension alone, it is able to predict the relative order of difficulty of trials on inclusive and exclusive disjunctions. Recent work on the analysis of the limitations of inductive learning algorithms (Valiant, 1984; Dietterich, 1989) is in sharp contrast to the versatility demonstrated by human learners. We believe that approaches that make use of background knowledge to focus all aspects of learning are central to accounting for the generality of human learning. Acknowledgements The software used to run the experiment was designed by Francis Nguyen and Takeshi Tsubota. We would like to thank Kamal Ali, Cliff Brunk, David Foster, and Scott Truesdel for assistance in running the experiment and Gupi Silverstein and Caroline Ehrlich for commenting on an earlier draft of the paper.