Neural Network Ising Machines: Algorithm Unrolling for Combinatorial Optimization

ICLR 2026 Conference Submission14252 Authors

Keywords: Ising Machines, Combinatorial Optimization, Algorithm Unrolling, Zeroth Order Optimization

TL;DR: We propose a new data-driven neural approach to combinatorial optimization in which we learn the parameters of an iterative dynamical system which efficiently solves typical instances of the NP-hard Max-Cut/Ising problem.

Abstract: We propose a new data-driven neural approach to combinatorial optimization in which we learn the parameters of an iterative dynamical system which efficiently solves typical instances of the NP-hard Max-Cut/Ising problem. The dynamical system is parameterized by a small neural network which is trained using a zeroth-order optimization method. We find that our method is able to learn efficient and scalable algorithms for solving these combinatorial optimization problems. We show that even with a limited parameter count, the neural network is able to learn sophisticated dynamics which allow it to efficiently navigate the non-convex landscapes that are characteristic of NP-hard problems. We compare our method against state-of-the-art neural-CO approaches as well as other classical Max-Cut/Ising solvers and show that is can achieve competitive performance.

Primary Area: optimization

Submission Number: 14252