Design of a Learning Machine and the Study of Some of Its Convergence Characteristics

摘要

Whenever the dynamics and environment of a process are unknown or very complex, there is a need for 'learning machines' capable of learning the optimal decision algorithm from experience. This dissertation proposes such a learning machine. The basic learning situation is specified by a set of six postulates and the machine MAXINE is developed to learn in this situation. MAXINE is designed to have some of the qualities of human decision making: while being able to 'change its mind' in the face of new evidence, it is reluctant to alter firmly held opinions. The learning ability of this machine is tested by placing it in situations of varying degrees of complexity, including those which are deterministic and stochastic. Convergence of the proposed learning algorithm for the deterministic case is proved. (Author)