We would like to invite you to a research seminar organised by the Aston Centre for Artificial Intelligence Research and Applications (ACAIRA). It will take place on the 19th of March at 13:00. Prof Michael Wong (Hong Kong University of Science and Technology) will deliver a talk entitled "Statistical Mechanics of Physics-inspired Optimization".

 

Abstract:

Discrete optimization problems are ubiquitous in science and engineering. The Ising model has become a powerful bridge for modeling and analyzing these problems from the viewpoint of statistical physics. However, finding the system states corresponding to the optimal solution of the target problems remains computationally hard. In recent years, a number of physics-inspired hardware systems have been proposed to solve optimization problems. Examples included superconducting ring networks leveraging quantum annealing, optimal systems based on pulse lasers, coupled oscillator networks, as well as special purpose chip-implementations on conventional hardware. In this talk, I will focus on a class of physics-inspired optimizers based on quantum optics known as Coherent Ising Machines. By applying the replica method to a few versions of Ising machines aiming at optimizing the Sherrington-Kirkpatrick model of spin glasses, we discovered a rich set of phases, with the solution variable distributions shown in Fig.1. This places our discoveries on a firm theoretical basis compared with previous heuristics. We found that for Ising machines to produce optimal solutions, it is important to locate the system in the region where two specific phases, the gapless phase and the binary phase, coexist. This is useful for the design of Ising machines and in addition the choices of parameters and their dynamical steering. By carefully placing a digitization operation in the driving term, we proposed a superior type of Ising machine, referred to as digCIM and satisfies the TAP equation with its variable distribution to be isomorphic to the gapless-binary coexistence condition. Our results show that digCIM reaches the top performance relative to a number of commercial optimization solvers in a global quadratic unconstrained binary optimization challenge.