Gradient Population Optimization: A Tensorflow-Based Heterogeneous Non-Von-Neumann Paradigm for Large-Scale Search
John Persano, Said M. Mikki, Yahia M. M. Antar · IEEE Access · 2018
This paper presents a novel scalable algorithm, Gradient Population Optimization (GPO), which is specifically designed to optimize cost functions with extremely high dimensionality. GPO uses the Tensorflow platform, a non-von-Neumann computation model, which implements dataflow graphs on heterogeneous computing hardware (e.g., multi-core central processing unit, graphics processing unit (GPU), and field-programmable gate array) in order to perform massively parallel processing tasks on scalable platforms, such as the cloud. GPO is based on the combination of population-based dynamics with gradient-based determinism, in which a coupling term is introduced between the local and global corrections to the positions of population's agents' positions. The GPO exhibited excellent performance in most of the standard benchmark functions that were tested. In particular, GPO demonstrated superb scalability in solving largescale optimization problems using GPU-hardware-accelerated computing platform, positing the algorithm as an effective strategy for real-life massive scale problems, such as machine learning, data mining, and modeling wireless communication systems, such as 5G and massive MIMO.