| Abstract: | Much natural learning occurs by observation without explicit feedback or tutoring, yet few models of learning address this class of tasks. Further, many natural cases of observational learning are complex, and efficient learning seems to demand strategic learning procedures. The present work adopts a design perspective and asks what learning mechanisms would be both useful and feasible for natural induction. Work on closely related learning problems is briefly reviewed and a model for observational rule learning is proposed and simulated. The model extends learning mechanisms developed for explicit learning with feedback to less structured, observational tasks. In particular, the focused sampling mechanism, which is an extension of the attentional learning procedure developed by Zeaman and House (1963), is introduced. The operation of an attentional learning procedure is less clear when extended to learning without feedback, so a simulation was done to evaluate the performance of the model. A series of simulated experiments were run, comparing performance of the learning model with and without the focused sampling component. We evaluated whether and when focused sampling benefits observational learning, investigated the effects of different distributions of systematic and unsystematic features, and compared observational learning to learning with feedback. Results of the simulation suggest that focused sampling does benefit learning, that benefit increases with the complexity of the learning task, and that learning with and learning without feedback exhibit differences in how each is affected by changes in the learning problem. Suggestions about the relation to human data are offered. |
