Software as Learning . Validation and Maintenance Issues

We reconsider validation and maintenance characteristics of software systems under the analogy between software science and philosophy of science or, more precisely, between software construction and machine learning (ML). From this outset, many classical techniques from ML can be used. In particular, we adapt a constructive extension of reinforcement learning to address the question in a formal way. We define a measure of software ‘predictiveness’, which is identified with software validation, to represent the stability of a system. An inversely related measure, the probability of modification, is also obtained for each component and for the whole system. The application in practice of these measurements is discussed. From here, we present some models of maintenance cost based on a detailed combination of predictiveness and modifiability. We study different software arrangement topologies theoretically. Hierarchised topologies, especially downward confluent ones such as trees and lattices involve less maintenance costs. Moreover, some intuitive rationales are confirmed, namely that compressed systems and coherent models (without patches or exceptions) are manifestly more maintainable.