Learning Local Search with Theoretical Indicators for Job Shop Scheduling

ICLR 2026 Conference Submission13032 Authors

Keywords: Combinatorial Optimization, Scheduling, Theoretical Indicator, Local Search

TL;DR: We integrated novel scheduling-specific properties as indicators into the action evaluation, thereby enhancing a learning-based local search method for job shop scheduling.

Abstract: Job shop scheduling problem (JSSP), where job sequences must be assigned across multiple machines to minimize makespan under fixed routes and varying processing times, is one of the most challenging combinatorial optimization problems. To improve search efficiency, we propose LSI, Local Search with Indicators, a learning-based local search method for JSSP. LSI integrates scheduling-theoretic conditions as indicators into the action evaluation, enabling the policy to focus on swaps that guarantee makespan reduction. By incorporating theoretically proven conditions into the action evaluation, LSI prioritizes promising swaps rather than treating all moves equally, representing a principled improvement of makespan. Despite relying only on a lightweight multilayer perceptron (MLP) policy network, LSI achieves competitive or superior performance compared to strong state-of-the-art approaches on diverse JSSP benchmarks, offering faster inference and robust scalability without retraining. These results demonstrate the effectiveness of embedding problem-structured theoretical principles into learning-based combinatorial optimization.

Primary Area: optimization

Submission Number: 13032