Improving Feasibility via Fast Autoencoder-Based Projections

2026-04-03

AI Summary

This work addresses the challenge of efficiently satisfying complex non-convex operational constraints in real-world learning and control systems. To this end, we propose a data-driven amortized projection method that leverages an adversarially trained autoencoder to learn a structured convex latent representation of the feasible set. Infeasible predictions are rapidly corrected by projecting them onto this latent space and subsequently decoding back to the original domain. By reformulating non-convex constraint enforcement as a convex projection problem in the latent space, our approach ensures constraint satisfaction while substantially reducing computational overhead. Experimental results across diverse optimization and reinforcement learning tasks with intricate non-convex constraints demonstrate that the proposed method significantly outperforms conventional solver-based correction strategies in both accuracy and efficiency.

Abstract

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.

Problem

• nonconvex constraints
• feasibility enforcement
• operational constraints
• constrained optimization
• reinforcement learning

Innovation

• autoencoder-based projection
• amortized feasibility correction
• convex latent representation
• nonconvex constraints
• adversarial training