DRL Transformer Solves Open Shop Scheduling with Scalability

This research presents a novel approach to tackle the computationally intensive Open Shop Scheduling Problem (OSSP), which is prevalent in various industrial and service sectors. The proposed method utilizes a Transformer-based deep reinforcement learning (DRL) policy, employing an encoder-decoder architecture with multi-head attention. The model was trained using only processing-time matrices on relatively small Taillard benchmark instances. A key finding was its ability to generalize effectively to much larger, unseen problems, ranging from 40x40 to 100x100 instances, without requiring retraining. When compared against traditional dispatching heuristics like SPT, LPT, MWKR, and EST, the Transformer achieved makespans typically within 15-30% of best-known values on small instances and maintained competitive performance on large instances. This demonstrates that a learning-based, feature-light Transformer policy can offer a viable and scalable alternative to classical scheduling rules, particularly for complex, large-scale OSSP scenarios.