Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

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.

Cong Zhang, Wen Song, Zhiguang Cao, Jie Zhang, Puay Siew Tan, Chi Xu• 2020

Related benchmarks

TaskDatasetResultRank
Job-Shop Scheduling ProblemDMU benchmark of JSSP
Average Gap (Instance-wise)35.4
104
Job Shop Schedulingta
Gap to BKS13.6
72
Job Shop SchedulingTaillard's benchmark Avg
Performance Gap (PG)16.41
57
Job Shop SchedulingTaillard Benchmark
Average Makespan1.55e+3
48
Job Shop SchedulingSynthetic 100 Instances per Size (val)
Average Makespan574.1
30
Job-Shop Scheduling ProblemTaillard 30 x 15
Optimality Gap (%)33
26
Job-Shop Scheduling ProblemJSSP 100 instances 10x10 (test)
Objective Value871.7
19
Job Shop SchedulingDemirkol's benchmark Avg
PG38.7
17
Job Shop SchedulingLawrence's benchmark Avg 40 instances
Average PG17.4
17
Job-Shop Scheduling ProblemTaillard 15 x 15
Optimality Gap26
17
Showing 10 of 45 rows

Other info

Follow for update