Abstract: This paper concerns a concept of selfperfecting and selflearning of digital computer systems. This idea is not new, but continuously in the state of elaborations. Such systems, i.e. with a possibility of selflearning and selfmodification would have undoubtedly greater possibilities and elastic properties in their behaviour than traditional digital systems. In the paper a digital system is treated as a complex system of algorithms. As an abstract model of real algorithms. so called Mazurkiewicz FC-algorithm is considered. FC-algorithms used in the paper have been extended to modifiable Fe-algorithms by adding a time-variant structure and the use of the notion of tolerance spaces. This structure allowed us to introduce a model of learning for modifiable FC-algorithms. Learning is understood as a goal directed process of changes of activities on the basis of experience.
Keywords: FC-algorithm – A, a modidiable FC-algorithm – MA, the ith configuration of MA – MAi, the ith function of modification – Φi, a relation of tolerance – τP, canonical sets of equations for MAi and for MA, a cost function – Ψ, an extended cost function – Ψ̅, a cost system – R, “beyond-real” numbers – R∞η, a finite learning process, a goal of learning, an optimization problem for MA, the stop-property for MA, an antigradient of the tolerance τP – ∇̅τP, a direction of the improvement, the outsider- and the inside-cost functions, a process of learning with a teacher and selflearning
