Learning to Dispatch for Job Shop Scheduling via Deep Reinforcement Learning
About
Priority dispatching rule (PDR) is widely used for solving real-world Job-shop scheduling problem (JSSP). However, the design of effective PDRs is a tedious task, requiring a myriad of specialized knowledge and often delivering limited performance. In this paper, we propose to automatically learn PDRs via an end-to-end deep reinforcement learning agent. We exploit the disjunctive graph representation of JSSP, and propose a Graph Neural Network based scheme to embed the states encountered during solving. The resulting policy network is size-agnostic, effectively enabling generalization on large-scale instances. Experiments show that the agent can learn high-quality PDRs from scratch with elementary raw features, and demonstrates strong performance against the best existing PDRs. The learned policies also perform well on much larger instances that are unseen in training.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Job-Shop Scheduling Problem | DMU benchmark of JSSP | Average Gap (Instance-wise)35.4 | 104 | |
| Job Shop Scheduling | ta | Gap to BKS13.6 | 72 | |
| Job Shop Scheduling | Taillard's benchmark Avg | Performance Gap (PG)16.41 | 57 | |
| Job Shop Scheduling | Taillard Benchmark | Average Makespan1.55e+3 | 48 | |
| Job Shop Scheduling | Synthetic 100 Instances per Size (val) | Average Makespan574.1 | 30 | |
| Job-Shop Scheduling Problem | Taillard 30 x 15 | Optimality Gap (%)33 | 26 | |
| Job-Shop Scheduling Problem | JSSP 100 instances 10x10 (test) | Objective Value871.7 | 19 | |
| Job Shop Scheduling | Demirkol's benchmark Avg | PG38.7 | 17 | |
| Job Shop Scheduling | Lawrence's benchmark Avg 40 instances | Average PG17.4 | 17 | |
| Job-Shop Scheduling Problem | Taillard 15 x 15 | Optimality Gap26 | 17 |