GOAL: New diffusion solver beats baselines by 25x on scheduling benchmarks
100% solution feasibility and near-zero error on multi-objective scheduling problems
Existing neural combinatorial optimization solvers are limited to single-objective static problems. GOAL (Graph-based Objective-Aligned Diffusion Solvers) breaks this mold by using a conditioned diffusion process over relational graphs, where human-specified objectives guide solution generation. A key innovation is heterogeneous graph encoding: different edge types (for different constraint classes) define selective message passing in the GNN, enabling the model to respect constraint ontology.
Evaluated on three canonical scheduling benchmarks (FSP, JSP, FJSP) with up to 20 jobs and 60 operations, GOAL achieves 100% feasible solutions and near-zero Mean Absolute Percentage Error (<0.20%) on multiple objectives. It generalizes zero-shot across different constraint regimes and problem types without architectural changes. Compared to NSGA-II and MOEA/D, GOAL delivers up to 25x faster inference and superior solution quality, making it a strong candidate for real-time dynamic scheduling in manufacturing, logistics, and resource allocation.
- Uses conditioned diffusion over graph representations with heterogeneous edge types to encode constraint ontology
- Achieves 100% solution feasibility and <0.20% MAPE on Flow Shop, Job Shop, and Flexible Job Shop problems
- Outperforms NSGA-II and MOEA/D by up to 25x in inference speed while improving solution quality
Why It Matters
Enables real-time, human-guided multi-objective optimization in dynamic scheduling, dramatically outperforming classical evolutionary methods.