Screening dense and noisy DOX-datasets with NN-blending and “dizzy” swarm intelligence: Profiling a water quality process

详细信息    查看全文

• 作者：George J. Besseris^a ; ^b ; ^{besseris@teipir.gr}
• 关键词：Stochastic optimization ; Taguchi methods ; Swarm intelligence ; GRNN ; Smart sampling ; Water quality
• 刊名：Swarm and Evolutionary Computation
• 出版年：2017
• 出版时间：February 2017
• 年：2017
• 卷：32
• 期：Complete
• 页码：219-233
• 全文大小：2178 K
• 卷排序：32

文摘

A novel nature-inspired method is presented in this work for resolving product/process development or improvement with design of experiments (DOX). The technique is suitable for difficult Taguchi-type multifactorial screening and optimization studies that need to simultaneously contain the double hassle of controllable and uncontrollable noise intrusions. The three-part sequential processing routine requires: 1) a regressive data-compression preprocessing, 2) a smart-sample generation using general-regression neural networks (GRNN), and 3) a screening power prediction using ‘reverse’ swarm intelligence (SI). The approach is primed to confront potential non-linearity and data messiness in the examined effects. The Taguchi-type orthogonal-array (OA) sampler is tuned for retrieving information in controllable (outer OA) and uncontrollable (inner OA) noises. The OA-saturation condition is elicited for maximum data exploitation. GRNN-fuzzification consolidates into a single contribution the uncertainty from all possible sources. The resulting ‘smart’ sample is defuzzified by a robust-and-agile data reduction. Screening-solution meta-power is controlled with a new SI-variant. The independent swarm groups, as many as the studied effects, are tracked toward preassigned targets, i.e. their ability to return to their host beehives. The technique is illustrated on a complex purification process where published multifactorial data had been collected for a critical wastewater paradigm and thus may be used to test the benchmark solution. However, environmental water-qualimetrics are profoundly dominated by messy data as justified in this work. We elucidate on several issues that regular Taguchi methods may be benefited by the proposed GRNN/SI processing while emphasizing the consequence of overlooking the underlying assumptions that govern standard comparison models. The new swarm itelligence method offered a practical way to estimate a first-time “soft” power measure for the inner/outer OA optimization case that was impossible with ordinary statistical multi-factorial treatments. Key performance advantages in efficiency, robustness and convenience are highlighted against alternative approaches.