SnareNet: Flexible Repair Layers for Neural Networks with Hard Constraints

Researchers have introduced SnareNet, a novel feasibility-controlled architecture designed to ensure neural network outputs strictly adhere to physical, operational, or safety constraints. While neural networks are increasingly utilized as fast surrogate models, their unconstrained predictions often violate critical requirements. SnareNet addresses this by appending a differentiable repair layer that navigates the constraint map's range space, steering predictions toward feasibility within a user-specified tolerance. The system employs a new training paradigm called adaptive relaxation, which stabilizes end-to-end training by initially allowing exploration and gradually shrinking the network into the feasible set. Benchmark tests in optimization learning and trajectory planning demonstrate that SnareNet consistently achieves improved objective quality and more reliable constraint satisfaction compared to prior methods. Notably, it is the first approach to robustly enforce non-convex constraints at medium-to-high precision across various instances. This advancement significantly enhances the reliability of AI models in safety-critical applications, marking a substantial step forward in constrained machine learning architectures.