Stochastic Gradient-Augmented Neuro-Symbolic Architecture for Constrained Multi-Objective Optimization in Autonomous Engineering Decision Systems

Abstract

Contemporary autonomous engineering systems demand decision-making frameworks capable of reconciling competing optimization objectives under dynamic, partially observable environments. This paper proposes a stochastic gradient-augmented neuro-symbolic architecture (SG-NSA) that integrates differentiable symbolic reasoning modules with deep reinforcement learning pipelines to address constrained multi-objective optimization in real-time industrial control scenarios. The proposed framework employs a dual-stream policy network coupled with a Lagrangian constraint satisfaction layer, enabling simultaneous maximization of operational throughput and safety compliance. Empirical validation across three benchmark engineering control environments demonstrates a 23.7% improvement in Pareto-front approximation quality over state-of-the-art baselines. Ablation studies confirm the critical contribution of symbolic grounding to out-of-distribution generalization. The architecture is shown to be computationally tractable for deployment on edge AI inference hardware, advancing the practical viability of intelligent autonomous engineering systems.