Neural-Network Based Self-Initializing Algorithm for Multi-Parameter Optimization of High-Speed ADCs

Abstract

This brief proposes a new automatic model parameter selection approach for determining the optimal configuration of high-speed analog-to-digital converters (ADCs) using a combination of particle swarm optimization (PSO) and stochastic gradient descent (SGD) algorithm. The proposed hybrid method first initializes the PSO algorithm to search for optimal neural-network configuration via the particles moving in finite search space with coarse quantization. Using the PSO estimates, the SGD algorithm then finds the global optimum solution. The global search ability of the PSO algorithm and the local search ability of the SGD are thus exploited to determine an optimal solution that is close to the global optimum with reduced latency. Several experiments were constructed to optimize the non-linearities in Nyquist flash and pipeline ADC datasets to show that the neural networks trained by the PSO-SGD algorithm outperform the random search-based performance optimization. Comparative resource analysis of the proposed algorithm is also conducted against the state-of-the-art that highlights improved latencies and performance with similar area and implementation complexity.