The Flexible Job Shop Scheduling Problem (FJSP) is a complex, NP-hard optimization challenge with significant practical relevance in manufacturing systems. This paper introduces a novel neural-driven constructive heuristic that replaces static priority dispatching rules with a compact, feed-forward neural network. The network evaluates potential operation-machine assignments during the schedule construction process based on features derived from the current partial schedule and job state. To train the network, we employ a black-box optimization approach using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES), eliminating the need for labeled data or reward functions. We propose a two-stage framework: general training across a broad range of synthetic or benchmark instances, followed by instance-specific fine-tuning of the network weights to adapt the heuristic for individual problem scenarios. On the Brandimarte benchmarks, our approach yields an average optimality gap of 0.91% (0.32% on the full dataset), outperforming the best deep reinforcement learning methods (2.35%) and the leading traditional dispatching heuristics individually suited to each instance of the problem (9.25%). Although CPLEX constrained programming solver achieves a slightly lower average gap of 0.05%, our method delivers competitive makespans overall and demonstrates superiority on numerous instances at six times lower computational cost. The results highlight the potential of learning-based constructive heuristics as a scalable and adaptable alternative for complex scheduling tasks. • Novel neural-driven heuristic for flexible job shop scheduling (FJSP). • A population of compact neural networks trained using CMA-ES. • A two-stage framework combining general training with instance-specific fine-tuning. • The heuristic achieves near-optimal makespans on Brandimarte benchmark tests. • Demonstrates advantages of population-based shallow learning over DRL in FJSP.
Kaleta et al. (2026) studied this question.