Reinforcement-Guided Neural Optimization for Low-Power Real-Time Analytics

Abstract

Real-time analytics on low-power devices requires neural models that operate efficiently under tight computational and energy constraints. Reinforcement learning offers a promising avenue for dynamically optimizing such models by enabling adaptive selection of execution paths based on observed sys tem conditions. This paper investigates a reinforcement-guided optimization framework designed to improve the efficiency of lightweight neural architectures deployed on constrained embedded platforms. The framework integrates structural adap tation, operator-level selection mechanisms, and reward-driven pruning strategies to balance inference accuracy with runtime cost. Experimental results demonstrate that reinforcement-guided optimization consistently improves throughput, reduces energy consumption, and stabilizes latency during continuous analytic workloads.