Back to results list
Please use this identifier to cite or link to this item:
|Title:||Towards capturing aesthetic intent of design in an interactive evolutionary system using neural networks||Authors:||Gu, Zhenyu||Keywords:||Hong Kong Polytechnic University -- Dissertations
Design -- Data processing
Neural networks (Computer science)
Evolutionary programming (Computer science)
|Issue Date:||2006||Publisher:||The Hong Kong Polytechnic University||Abstract:||This thesis has developed an interactive system that uses an evolutionary algorithm incorporating an artificial neural network for improving the aesthetic appeals of design. The artificial neural network allows the system to generalise user preferences by learning the implicit relationship between the evolved design solutions and the user evaluations. Evolutionary algorithms are usually used as optimizers for searching the best solutions to specific problems by imitating the adaptive processes in nature. These algorithms must, however, rely on predefined objective functions called fitness functions, in order for the system to converge on the optimal solution. The formulation of the fitness functions is a major bottleneck in the application of these algorithms in design domains. In recent years, evolutionary algorithms have been used in computer aided design systems for generating desirable images or 3D forms. These systems use human subjective evaluations and selections instead of objective fitness functions to control the evolution, and they are referred to as the Interactive Evolutionary Systems (IES). An IES is capable of generating and evolving large numbers of alternative designs as well as finding the optimum design. In an application, a designer is continuously required to interact with the system by making evaluations and selections of the designs that are being generated and displayed on a computer screen. However, the IES approach involves a process which may become unendurable and exhaustive for several reasons. First the limited size of a computer screen allows only a small number of candidate solutions to be displayed, evaluated and selected by the designer. Second, due to fuzzy nature of aesthetic evaluations, evolution is usually a mutation-driven and divergent process. Third the evolutionary process is time consuming due to limited speed of ratings and interactions. The convergent mechanisms typically found in standard Evolutionary Algorithms are more difficult to achieve with IES. In an IES, optimisation gives way to exploration due to the open ended nature of the algorithms.
To address this problem, the thesis proposes an approach to use a neural network in conjunction with an IES to obtain a smoother, less erratic evaluation function than what would be the direct result of user's individual choices. A designer comes to interact with the evolutionary system and makes evaluations and selections. If the designer's aesthetic preference is consistent, the seemingly erratic user selections may have certain consistency and tendency. The designer's selections are recorded and the artificial neural network is thereby able to approximate implicit correlation between evolving designs and the designer's responses. This allows a canonical evolutionary process to run with much larger population samples to get a good distillation of designs that the designer most likely prefers. This process is called "capturing user intent". The approach supports a design paradigm which consists of iterative cycles of interactive evolution and intent capturing, as the intent of the user may not be exactly captured in a single try. The neural network for the approximation is a kind of RBF network called General Regression Neural Network (GRNN). The approximation is a regression of aesthetic appeals conditioned on the corresponding designs. The basic idea for this framework is based on the assumption that learning behaviours exist in design processes. The inductive learning ability of human beings can generate and adapt designs based on past failures and successes, whilst in natural evolution the generation, reproduction and variation are blind processes. Therefore a learning mechanism in an evolutionary design process contributes to the formulation of aesthetic intent of a designer in terms of an approximated fitness function for shortening the tedious and lengthy process of human evaluation and selection involved in an IES. In such an approach, a balance is reached between exploration and optimisation, both of which are essential tasks of design and should be supported by computational power. The application for the implemented system focuses on parametric tuning activities of design. A prototype system for facial character creation is implemented in order to study the feasibility of the proposed new framework. The implemented system is interactive and it uses genetic algorithms for searching desirable 3D facial surface models. An intuitive interface with easy to use operations provides a basis for testing and evaluating the implemented system. Several experiments are conducted in order to verify the performances of the system implemented. The outcomes are analyzed and evaluated.
|Description:||161 leaves : ill. ; 31 cm.
PolyU Library Call No.: [THS] LG51 .H577P SD 2006 Gu
|URI:||http://hdl.handle.net/10397/3626||Rights:||All rights reserved.|
|Appears in Collections:||Thesis|
Show full item record
Files in This Item:
|b20938664_link.htm||For PolyU Users||161 B||HTML||View/Open|
|b20938664_ir.pdf||For All Users (Non-printable)||5.53 MB||Adobe PDF||View/Open|
Citations as of Jun 18, 2018
Citations as of Jun 18, 2018
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.