PLAN Boosts Job Shop Scheduling Efficiency with Parallel Liquid Networks

Deep reinforcement learning (DRL) has achieved state-of-the-art results in flexible job shop scheduling (FJSP), but current attention-based models are often too large and slow for practical, large-scale applications. This research proposes PLAN, a Parallel Liquid-inspired Approximation Network, which tackles these limitations by re-imagining liquid neural network (LNN) dynamics into a parallelizable, discrete form. PLAN separates the evolving state representation, handled by liquid-inspired updates, from global context aggregation, managed by a lightweight module. PLAN acts as a flexible backbone, adaptable to various FJSP complexities, including stochastic and multi-faceted dynamic scenarios. Extensive testing across these benchmarks shows PLAN reduces average makespan by 1.2% to 2.3% compared to leading baselines, with some improvements reaching over 10%. Crucially, it also slashes average inference latency by 13.2% to 31.7%, and up to 69.2% on larger instances, all while using only 22-47% of the parameters of its counterparts.