基于线性变换的适应度函数及机器人进化计算研究

【作者】 彭伟民；

【导师】 唐平；

【作者基本信息】 广东工业大学 ， 计算机应用技术， 2006， 硕士

【摘要】 智能化是计算机机发展的必然趋势,无论是计算机控制,还是商用民用软件,都要求含有越来越高的智能因素,因此人工智能的研究越来越受到重视。20世纪80年代,基于结构演化的人工智能理论——计算智能理论迅速成为人工智能研究新的主流。计算智能包含广泛的研究领域,各领域之间存在着深刻的联系,且相互促进,进化计算就是其中一个重要领域。 进化机器人的思想主要来源于进化计算。在进化机器人中,设计者的工作主要是决定进化框架和评估策略,评估中所采用的适应度函数在很大程度上决定了系统的行为。常用的适应度函数的设计方法有线性变换法、幂函数变换法、指数变换法和方差调整法等。同时,进化框架中的进化参数,如选择概率、交叉概率和变异概率等对进化过程和结果非常关键。通过研究机器人进化计算中适应度函数及进化参数的设计,达到群体多样性与收敛性、进化性能与进化速度的统一,从而优化进化机器人中进化框架和评估策略的设计。这对于进化计算和进化机器人学的发展都具有重要的意义。 本论文研究内容:运用线性变换法研究机器人进化计算中的适应度函数的选择及进化参数的改进,着重研究适应度函数的选择。在线性变换法中,比率系数和幅度系数最终决定线性变换的结果,而幅度系数由比率系数和适应度平均值系数决定。根据不同的进化阶段,建立比率系数和适应度平均值系数自适应变换的表达式,得到按比率、幅度及比率和幅度三种基于线性变换的适应度函数的进化计算。而进化参数表达式的建立,则需要根据进化阶段的不同和个体适应度的大小而定。 本文创新之处:通过建立根据进化前期或后期自适应变换的比率系数、适应度平均值系数及选择概率、交叉概率和变异概率等进化参数表达式,构造了三种基于线性变换的适应度函数及进化参数的数学模型,设计相应算法,然后在Evorobot系统中进行了相关的仿真实验,并对仿更多还原

【Abstract】 Now the computer is tending to more and more intelligent, and no matter what kind of applications, intelligence has been the most important factor. In the 1980s, the artificial intelligence theory based on the evolution of structure — computational intelligence was fast becoming the new mainstream. Computational intelligence consists of a wide range of research areas which have profound links and promote each other, and evolutionary computation is just an important field of these areas.The thinking of evolutionary robotics mainly comes from evolutionary computation. In evolutionary robotics, the primary work of designers is to decide evolutionary framework and assessment strategies. The system behaviors depend to a high degree on the fitness function used in assessing. In designing methods of fitness function there are many ways, for examples: linear transformation, power transformation, index transformation, variance adjustment, etc. Meanwhile, the parameters in evolutionary framework, such as the selection probability, crossover probability and mutation probability are crucial to the evolutionary process and results. The target of this paper is to achieve the integration of the populations’ diversity and convergence, evolution speed and performance, thereby optimize the design of evolutionary framework and assessment strategies, with the design of fitness function and evolutionary parameters. This is of great significance to the development of evolutionary robotics and evolutionary computation.This paper includes the following studies: research of thechoices of fitness function and the improvement of evolutionary parameters with linear transformation, focusing on the choices of fitness function. In linear transformation, the ratio coefficient and the scope coefficient ultimately determine the outcomes, and the scope coefficient is determined by the ratio coefficient and the fitness average coefficient. Depending on different evolutionary stages, we will establish the self-adapting expressions of the ratio coefficient and the fitness average coefficient, and then obtain three kinds of evolutionary computation based on linear transformation which have different fitness functions according to ratio, scope, ratio and scope. And the construction of evolutionary parameters depends on different stages of evolution and the size of fitness.The innovations are included in this study. The self-adapting expressions of ratio coefficient, fitness average coefficient and evolutionary parameters which alternate by the early or late stage of evolution are established, and then the mathematical models for three fitness functions and evolutionary parameters based on linear transformation are proposed and the corresponding algorithms are designed. After that, the related simulation experiments are implemented on platform Evorobot, then the simulation results are analyzed and the corresponding conclusions are got.更多还原