Constitutional design is a new grammatical relationship and thinking mode of modern morphological design. It reflects the basic characteristics of modern design that engages in the essence of things and uses analysis and comprehensive scientific methods to create shapes. Genetic algorithms are bionic algorithms in a macro sense and a branch of evolutionary algorithms. The structure of neural networks considerably affects their performance. If there are too many neurons in a network, it may fit the training data well, and the network may have few training errors, but it can be prone to problems such as “overfitting” and poor network generalization. Based on the this background, in this article, the research content is based on genetic algorithms to optimize neural networks’ indoor space composition. This paper starts with genetic algorithms and proposes an improved genetic algorithm based on the neural network structure optimization algorithm (IGA). Based on the proposed algorithm, MATLAB is used as an experimental platform to approximate the non-linear function $y=e^{-{(x-1)}^2}+e^{-{(x+1)}^2}$. The experiment proves that the IGA has a certain optimization effect on the composition of the neural network structure in indoor space. The best individual fitness value is 0.786, which improves the neural network’s adaptability and generalization, providing good global rapid convergence performance.

宪法设计是现代形态设计的一种新的语法关系和思维方式。它反映了参与事物本质的现代设计的基本特征，并使用分析和全面的科学方法来创建形状。遗传算法是宏观意义上的仿生算法，是进化算法的一个分支。神经网络的结构极大地影响了它们的性能。如果网络中神经元太多，则可能很好地适合训练数据，并且网络可能没有训练错误，但它容易出现诸如“过度拟合”和网络泛化不佳的问题。在此背景下，本文的研究内容基于遗传算法，以优化神经网络的室内空间组成。本文从遗传算法入手，提出了一种基于神经网络结构优化算法（IGA）的改进遗传算法。基于所提出的算法，将MATLAB作为近似非线性函数的实验平台$ÿ={Ë}^{-{（X-\mathrm{1个}）}^2}+{Ë}^{-{（X+\mathrm{1个}）}^2}$。实验证明，IGA对室内神经网络结构的组成具有一定的优化作用。最佳个体适应度值为0.786，可提高神经网络的适应性和泛化能力，并具有良好的全局快速收敛性能。