TY - JOUR
T1 - Experimental Validation of Configurational AI Architectures
AU - Yavich, Roman
AU - Rotkin, Vladimir
N1 - Publisher Copyright:
© 2026 by the authors.
PY - 2026/9
Y1 - 2026/9
N2 - Featured Application: The reproducible evaluation protocol and the procedural hybrid pattern examined here (a neural network generates a candidate configuration; an independent verifier checks all hard constraints; and an exact deterministic solver is invoked only on verification failure) are intended to inform the design of configurator systems in education, engineering, and e-commerce, in which recommendations must satisfy mandatory budget, quality, safety, and compatibility rules, and in which an auditable log of the accepted route is required for compliance and quality assurance. The evidence reported below was obtained on a synthetic corpus and therefore supports the architectural pattern itself; deployment in any of these settings would require separate validation on domain data. Configurable intelligent design is formalized here as the selection of a consistent configuration from interdependent alternatives under constraints of budget, quality, risk, and compatibility. It is examined not as a universal solver but as a class of domain-specific systems, each defined for a task family with explicitly formalizable variables, hard constraints, and objectives. Locally correct predictions for individual components do not guarantee a globally feasible solution. This study compares direct neural-network inference, the exact HIM-D solver, and a procedural hybrid whose network candidate is checked by an independent verifier, with an exact search invoked whenever a constraint is violated. A reproducible synthetic corpus of 24,000 problems across educational, engineering, and commercial scenarios was assessed for feasibility, optimality, robustness to distribution shift, repeatability, and latency. Pure neural-network models were feasible in only 17.40% and 19.12% of cases and degraded sharply out of distribution, whereas the hybrid maintained 100.00% feasibility and 83.08% joint optimality. A staged ablation identifies pairwise incompatibility density as the dominant source of that degradation, and constraint-aware correction alone raises feasibility without reaching a guarantee. The results therefore support separating probabilistic candidate generation from independent formal verification within a bounded class of specialized configuration problems, not a universal solver.
AB - Featured Application: The reproducible evaluation protocol and the procedural hybrid pattern examined here (a neural network generates a candidate configuration; an independent verifier checks all hard constraints; and an exact deterministic solver is invoked only on verification failure) are intended to inform the design of configurator systems in education, engineering, and e-commerce, in which recommendations must satisfy mandatory budget, quality, safety, and compatibility rules, and in which an auditable log of the accepted route is required for compliance and quality assurance. The evidence reported below was obtained on a synthetic corpus and therefore supports the architectural pattern itself; deployment in any of these settings would require separate validation on domain data. Configurable intelligent design is formalized here as the selection of a consistent configuration from interdependent alternatives under constraints of budget, quality, risk, and compatibility. It is examined not as a universal solver but as a class of domain-specific systems, each defined for a task family with explicitly formalizable variables, hard constraints, and objectives. Locally correct predictions for individual components do not guarantee a globally feasible solution. This study compares direct neural-network inference, the exact HIM-D solver, and a procedural hybrid whose network candidate is checked by an independent verifier, with an exact search invoked whenever a constraint is violated. A reproducible synthetic corpus of 24,000 problems across educational, engineering, and commercial scenarios was assessed for feasibility, optimality, robustness to distribution shift, repeatability, and latency. Pure neural-network models were feasible in only 17.40% and 19.12% of cases and degraded sharply out of distribution, whereas the hybrid maintained 100.00% feasibility and 83.08% joint optimality. A staged ablation identifies pairwise incompatibility density as the dominant source of that degradation, and constraint-aware correction alone raises feasibility without reaching a guarantee. The results therefore support separating probabilistic candidate generation from independent formal verification within a bounded class of specialized configuration problems, not a universal solver.
KW - ablation analysis
KW - combinatorial optimization
KW - configurational AI architectures
KW - constraint programming
KW - constraint satisfaction problems
KW - formal verification
KW - hierarchical imitation models
KW - hybrid artificial intelligence
KW - out-of-distribution generalization
KW - reproducible experiment
UR - https://www.scopus.com/pages/publications/105051663541
U2 - 10.3390/app16189265
DO - 10.3390/app16189265
M3 - ???researchoutput.researchoutputtypes.contributiontojournal.article???
AN - SCOPUS:105051663541
SN - 2076-3417
VL - 16
JO - Applied Sciences (Switzerland)
JF - Applied Sciences (Switzerland)
IS - 18
M1 - 9265
ER -