Experiments focused on (a) how a subject directs his attention and explores the problem environment created by the task to be learned, and (b) how the subject's learning process can be facilitated by CAI systems. In all experiments, the subject interacted with an automaton responsible for controlling the experimental environment. The automaton was either a procedural device that imposed behavioral constraints that represented rules of the task but permitted the subject freedom, particularly in his choice of a strategy, or it was a simple adaptive teaching machine or a conversational machine. The simulation incorporated subprograms representing the subject and automaton respectively. Different subjects exhibited different strategies for learning the same skill. Results show that an adaptive metasystem is a better instructional instrument for the skill studied than either the attention-directing process in a free learning subject or a more restricted, single-strategy teaching system. The computer simulation of the learning process in the subject is a psychologically oriented artificial intelligence program, not a stochastic model. It is heterarchic in that its structure involves several separable but interacting hierarchic systems. (Author/GO)