[20p-F211-1~13] 31.1 Focused Session "AI Electronics"

[20p-F211-10] Improvement of Simulated Bifurcation Algorithm by Thermal Fluctuations

〇Taro Kanao¹, Masaru Suzuki², Kosuke Tatsumura¹, Hayato Goto¹ (1.Toshiba R&D Center, 2.Toshiba Digital Solutions)

Keywords:combinatorial optimization problem, Hamiltonian system, annealing

Recently we have proposed an algorithm which solves large-scale combinatorial optimization problems by simulating adiabatic bifurcations in nonlinear Hamiltonian systems with fast digital parallel computers such as an FPGA and a GPU. In the present study, to this algorithm we introduce thermal fluctuations by using the Nose-Hoover method, and decrease the fluctuations with time, which is an annealing. As a result, we obtain solutions with improved accuracies. We think that this improvement is because the fluctuations enlarge search areas and the annealing improves convergence to better solutions.