Human Computer Cooperative Evolutionary Design Based On Distribution Estimation

The layout design of satellite module is one of the key issues in the global design of satellite,which has an important impact upon reducing the design period,saving the cost,improving the performance for the satellite.It is known as a hybrid combinatorial optimization and NP-hard problem in mathematics,and a scheme design and complex system problem in engineering.There exist two primary challenges in the layout problem, which are the combinatorial explosion in mathematics and the complexity in engineering system.A project of Research and Development of a Software Platform for Layout Design of Spacecraft from a research institute of China Aerospace Science and Technology corporation (CAST) is taken as a background.The work aims to improve the global mass properties of a satellite by adjusting the locations and orientations of given components. Moreover,it is supported by the National Nature Science Foundation of China.The main contributions of this work are as follows.1.A new PCA-EDA algorithm which combines Principle Component Analysis(PCA), a classical statistical technique,and Gaussian probabilistic model is given.EDAs are population-based search algorithms based on probabilistic modeling of promising solutions in combination with the simulation of the induced models to guide their searching. The probability factorization can be employed to encode the conditional(in)dependencies among different variables and overcome the linkage problem.But there are still some problems such as premature convergence and complex model learning which limit the class of problems that EDAs can solve reliably and efficiently.PCA-EDA is aimed to keep the balance of accuracy and efficiency of model learning,as well as to avoid premature convergence by easy variance control.2.The thesis explores the landscape of packing problem and supposes that it is characterized as the complex mixture of big-valley and symmetry structure.It has been shown that for any optimization algorithm to be successful when solving a certain problem, the structure of the problem needs to match the bias of the algorithm.So it is necessary to introduce landscape analysis method based clustering into PCA-EDA.Caussian Mixture Model(GMM) is used in the work which would improve the algorithm in three ways:(1) splitting the solution space and extracting the high performance region to help user to understand the problem and make the algorithm more efficient;(2) modeling the complex landscape of packing problem more accurately than single peak probabihty distribution; (3) equivalent to a EA with multiple populations that is help to keep diversity.To make the fusion of GMM and PCA-EDA more effectively,the thesis presents a multi-stage greedy clustering strategy inspired by the similarity of working manners of EA and Greedy EM.3.PCA-EDA's inner working way is analyzed from the point of view of design knowledge extraction.The design knowledge inducted by PCA-EDA is the variable correlation model which can be used to decompose the large-scale layout design problem into individual sub-problems in order to help designer to understand how the algorithm work.Then the relations between the variances of principle components,the correlation of original variables and the structure of landscape of the problem are studied in order to provide a more effective interaction manner from human to computer.Moreover,to help designers to capture key information during Human Computer Cooperation,four new interactive interfaces are presented after the survey of the state of software visualization for EAs. These interfaces can be used to gain insight into the state and the course of the algorithm and the landscape of the problem,which impact the performance of HCC heavily.Summarizing,the work focuses on variable correlation,visualizing and cooperation of qualitative analysis and quantitative analysis,and landscape structure.By considering and utilizing the characteristics of the EAs and the layout problem,the three aspects of the strategy of layout problem solving are unified closely to construct a human computer cooperative evolutionary design based on statistical learning techniques,e.g.distribution estimation.From the results of numerical experiments for the satellite module it is showed that the proposed method is feasible and effective for the layout design of satellite module. This work is expected to advance the theories of HCC evolutionary design for the complex system and layout design for satellite module in theory,and is also likely to benefit the research and development of practical methods and techniques for a layout design problem of satellite module.Finally,the main contributions of this research are likely to be applied to other complex layout problems.