Answers are grounded in this paper and cite their sources.
Numerical optimization in real time has proven to be a reliable and essential tool for applications such as energy consumption and production optimization, path planning in autonomous driving, model predictive control (MPC), motion planning for legged robots, constrained inverse kinematics (IKs), and traffic engineering. However, the high computational demands of numerical optimization methods often result in increased costs or delays that compromise efficiency and, in some cases, feasibility. To address these challenges, research efforts have focused on making numerical optimization algorithms more computationally efficient and less resource-intensive by leveraging suboptimal solutions, exploiting problem structure, and/or approximating solutions using artificial neural networks (ANNs). In particular, neural networks have demonstrated considerable success in solving MPC problems and continuous optimization problems, especially those involving inequality constraints. In this article, we establish sufficient conditions under which a neural network can approximate the solutions to a family of general continuous optimization problems with equality and inequality constraints. We also demonstrate how these conditions can be leveraged to streamline the data collection process, reducing the effort required to generate training datasets. To the best of our knowledge, this constitutes the most general approach of its kind in the literature. As a practical demonstration, we present a shared-load drone experiment to illustrate the use of neural networks for approximating the solutions of two optimization problems: one that computes the system's IKs and another that approximates the MPC solution in a soft real-time setting.
No takes yet. Share an insight, caveat, or question.
López-Rojas et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: