This paper describes CHREST+, a computational model which learns perceptual chunks to solve problems using a diagrammatic representation. Perceptual chunks are pieces of familiar information, as retrieved by a sensory device. In earlier work on chess expertise, a successful computational model, CHREST, has been developed of how such chunks can be acquired and stored in a discrimination network. CHREST+ is an extended version which learns associations between chunks for problem and solution states to create a knowledge base of information for problem solving. We compare the use of chunks by the model and human subjects in a problem-solving domain where unknown quantities are computed from electric circuits using a diagrammatic representation. We also discuss how the learning mechanisms of CHREST+ differ from those of ACT-R and Soar.
To receive an off print, please email Irene Jackson, or download the postscript version.