Neural Quantum Annealing for Real-World Quadratic Unconstrained Binary Optimization

We simulate Quantum Annealing on a variational manifold defined by a parametric family of wavefunctions represented by a Restricted Boltzmann Machine architecture. By iteratively lowering the transverse field and optimizing the neural network parameters, we prove the effectiveness of our methods for a challenging class of real-world Quadratic Unconstrained Binary Optimization (QUBO) instances, namely the Job Shop Scheduling Problem. Our methods outperform a straightforward optimization of the same problem without simulated quantum fluctuations. Despite the QUBO model being equivalent to a frustrated Ising spin glass, we obtain exact solutions up to 50-100 binary variables and empirically outperform results obtained with D-Wave quantum annealers.