This paper presents the concepts of a computer-directed system to improve human performance in structured learning situations. Attention is focused on finite-state systems in order to provide a systematic method for constructing training systems and to assist in analysis of problem solving and curriculum planning. The finite-state approach allows structural relationships among instructional units relative to individual problem solvers to be captured. At this time a fully adaptive system that improves its own performance with use requires further study and development, but it is possible for computer-directed systems that generate and select instructional actions to be used in both training and experimental situations. (CH)