NEURO-SYMBOLIC ARTIFICIAL INTELLIGENCE: INTEGRATING LOGICAL REASONING AND MACHINE LEARNING FOR ENHANCED DECISION-MAKING

Abstract

Traditional rule-based systems have transparent inference but fail with noisy, highdimensional inputs, while pure connectionist models predict well from raw data but are opaque function approximators. This dichotomy persists despite safety-critical decision-making requiring auditable rationale and reliable perception. NSDM, a neuro-symbolic decision-making system, connects a learned neural sensing module to a differentiable symbolic reasoning layer using confidence-aware fusion. A fuzzy first-order logic layer verifies domain rule satisfaction, a neural component encodes structured and unstructured inputs into a latent representation, and a gating network arbitrates between the two evidence streams to produce a calibrated decision and a humanreadable rule trace Composite goals penalise classification loss for semantic consistency, encouraging the network to predict from the encoded knowledge base. This framework is tested against a rulebased expert system, gradient-boosted trees, a deep neural baseline, and a novel neuro-symbolic technique on a consolidated high-stakes decision dataset. With less than 9 ms decision latency, NSDM achieves 93.6% accuracy, precision, and recall. Each conclusion is explained in detail by the framework, improving accountability, accuracy, and numerical improvements. Additionally, deployment, calibration, and implementation constraints were addressed.