Methods for constructing intelligent software and information decision support systems based on machine learning and knowledge management systems

Abstract

Multi-criteria assessment of regional environmental protection measures runs into a persistent contradiction: expert procedures scale poorly and depend on the composition of the working group, while rigid formal indices lose sensitivity as the set of measures changes. Machine learning lends the assessment adaptivity but returns an opaque result, whereas a knowledge base built on production rules and a domain ontology offers transparency yet does not adjust to the accumulating data. The aim of the work is a method for designing an intelligent decision support system in which the statistical and deductive components are coupled so that neither property is lost. The proposed four-layer architecture combines a predictive layer based on gradient boosting, a knowledge layer of production rules and an ontology of criteria, and an aggregation layer that reduces partial scores to an integral indicator over three criteria – ecological effectiveness, economic feasibility and long-term sustainability. On a model sample of 480 measures the hybrid configuration reached F1 = 0.891 against 0.847 for standalone boosting, reducing the share of decisions requiring manual revision from 0.34 to 0.11. A Monte-Carlo sensitivity analysis showed that the integral indicator tolerates a weight perturbation within +-0.07 without rank reversal for the top quartile of measures. Interpretability in such systems is attained by redistributing the explanatory and constraining functions to a coupled knowledge base rather than by abandoning machine learning; the design problem thereby shifts from the choice of algorithm to the organization of interaction between the statistical and logical modules.