The experimental design is critically important in the first stage of data analysis To collect data we defined the optimal criteria based on the experimental requirements and then generated the design accordingly An optimal design offers accurate inferences at minimal cost Theoretically proving a design's optimal structure is a typical approach to obtaining a proper design but there are often restrictions Such restriction can be on the run size of the optimal design and the structure of the statistic model Essentially we could treat design-construction problems as optimization problems and then use numerical algorithms to find the optimal design The exchange-type algorithm is a commonly used tool for design construction In practice we ran this algorithm simultaneously many times with various initial designs and then output the best one of those results as the exchange-type algorithm can easily become trapped in the local optimizers For this reason in this dissertation we propose a novel way to efficiently construct optimal designs This method involves a class of algorithms called swarm-intelligence algorithms which allow the initial starters to learn from each other through a specialized learning mechanism In this dissertation we propose two swarm-intelligence algorithms related to discrete design a two-level discrimination design for a set of linear models with main effects and two-factor interactions and an orthogonal array for estimating the main effects and some prespecified two-factor interactions In the end we introduce our novel development of design-generating software for practical use

Swarm Intelligence Algorithms for Design Generators

秉洋, 陳. (Author). 2018 一月 25

學生論文: Doctoral Thesis