Authors:

Charles Spencer Bowles

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

There exists a class of hard problems known as combinatorial optimization problems within the cores of logistics, finance, drug discovery, machine learning, and several other fields. These problems involve selecting the best configuration from a finite set of options. The broad consensus amongst computer scientists is that their unique properties prohibit conventional computers from finding solutions at useful problem sizes. Therefore, improved hardware architectures are needed to address these challenges. Recently, researchers are exploring a class of computational systems known as Coherent Ising Machines (CIMs) for their inherent ability to accelerate these problems. CIMs are optical computers that represent the elements of combinatorial optimization problems as pulses in a feedback system, naturally progressing to optimal solutions. Due to their analog nature, CIMs suffer from the phenomenon amplitude inhomogeneity, whereby nonbinary spins misrepresent a problem's underlying parameters, decreasing quality of results. We seek to alleviate this issue by investigating and combining previously proposed modifications to traditional CIMs. Modifications include new combinations of both network dynamics and spin couplings achievable in optical and electronic platforms. We provide an extensible benchmarking system capable of efficiently simulating broad classes of CIMs and assessing their performances and times-to-solutions on the standard BigMaq and G-set datasets. We report promising results for the use of multiple improvements on a traditional CIM. Uniform benchmarking helps provide insight to the relative efficacies of various promising CIM architectures prior to their physical construction, and we encourage further use of this benchmarking system through a publicly-available Python API. Keywords: Optimization; Unconventional Computing; Photonics; Simulation

Source:

Purdue University / 2024

Topics:

No topics listed

Co-authors:

Charles Spencer Bowles