Improving Feasibility via Fast Autoencoder-Based Projections
It offers a practical alternative to expensive solver-based feasibility correction for practitioners dealing with complex constraints in real-world systems.
The paper proposes a fast autoencoder-based method to enforce nonconvex operational constraints in learning and control systems, achieving low-cost feasibility correction across diverse constrained optimization and RL problems.
Enforcing complex (e.g., nonconvex) operational constraints is a critical challenge in real-world learning and control systems. However, existing methods struggle to efficiently enforce general classes of constraints. To address this, we propose a novel data-driven amortized approach that uses a trained autoencoder as an approximate projector to provide fast corrections to infeasible predictions. Specifically, we train an autoencoder using an adversarial objective to learn a structured, convex latent representation of the feasible set. This enables rapid correction of neural network outputs by projecting their associated latent representations onto a simple convex shape before decoding into the original feasible set. We test our approach on a diverse suite of constrained optimization and reinforcement learning problems with challenging nonconvex constraints. Results show that our method effectively enforces constraints at a low computational cost, offering a practical alternative to expensive feasibility correction techniques based on traditional solvers.