Abstract
<jats:p>This paper addresses the systematic design of neural network architectures for tabular data processing under stringent resource constraints typical for embedded and edge systems. The work aims to develop a methodology that formalizes this task as a constrained multi-objective optimization problem within a discrete hyperparameter space, framing the synthesis process as decision support under uncertainty. A methodology for multi-criteria evolutionary synthesis is proposed, based on the Multi-Island Genetic Algorithm (MIGA), which integrates an island model of evolution to maintain population diversity and the NSGA-II selection mechanism to construct a Pareto front approximation. Conflicting optimization criteria include classification accuracy, memory footprint required for model storage, and inference latency. For the experimental validation of the methodology, three public tabular datasets representing different application scenarios and complexity levels were selected. A software framework with a three-tier architecture was developed and implemented, supporting the full cycle of automated design—from adaptive data analysis to results visualization. A comparative analysis with baseline methods (logistic regression, decision tree, gradient boosting) demonstrated that the proposed methodology can synthesize models that, with comparable accuracy, are orders of magnitude more compact and faster than gradient boosting models. In cases involving complex nonlinear dependencies with small sample sizes, the synthesized models statistically significantly outperform the baselines in accuracy. The results confirm the practical significance of the methodology for reducing design complexity and providing developers with a set of quantitatively justified trade-off solutions that comply with given hardware constraints.</jats:p>