Quantum-enhanced neuro-symbolic intelligence for secure and adaptive autonomous systems : A computational optimization framework
Indra Kishor, Udit Mamodiya
Abstract
Autonomous intelligence needs to be predictive, constraint-satisfying, have calibrated uncertainty, and be efficient in decision computation. Neural approaches can be statistically successful, but also be sensitive to the variability of observations and/or expensive with respect to decision ambiguity. In this study, a Quantum-Enhanced Neuro-Symbolic Intelligence (QENSI) framework is proposed, which combines neural security inference, uncertainty calibration, symbolic feasibility constraints, adaptive quantum invocation, and cost-aware decision processing. The leakage-controlled benchmark, independent training, validation, calibration and test partitions, five pre-defined seeds, robustness perturbations, ablation studies and sensitivity analysis were used for evaluation. QENSI had an AUROC of 0.9445, an expected calibration error of 0.0193 and zero clean-test constraint violations with a macro-F1 of 0.7408. The selective quantum invocation was 52.5% which was significantly lower than the unconditional execution. Removing the quantum component decreased macro-F1 and AUROC, and removing symbolic reasoning increased the number of constraint violations. In the case of feature masking, the results were very similar, with 99.31% of clean macro-F1 remaining in the case of mild feature masking, but significant degradation in the case of severe cyber-risk escalation. The results suggest that the adaptive quantum-neuro-symbolic integration can enhance the quality of the balanced decision, calibration, and consistency of the constraints in the considered computational environment. Closed-loop planning, physical quantum hardware, and distribution-shift adaptation should be explored in the future for validation.