| | 尽管Koopman数据驱动算法前景广阔,但要选取合适的观测函数仍然是一个核心挑战。即使对于相对简单的动力系统,Koopman 算子的特征函数也可能任意复杂,并且在有限基中只能被近似地表示。深度学习非常适合表示这种任意复杂的函数,并且最近在Koopman 预测方面显示出了巨大的潜力<ref>C. Wehmeyer and F. Noe\', Time-lagged autoencoders: Deep learning of slow collective variables for molecular kinetics, J. Chem. Phys., 148 (2018), pp. 1--9. (Cited on pp. 284, 290, 291)</ref><ref>N. Takeishi, Y. Kawahara, and T. Yairi, Learning Koopman invariant subspaces for dynamic mode decomposition, in Advances in Neural Information Processing Systems, 2017, pp. 1130--1140. (Cited on pp. 284, 290)</ref>。除了利用神经网络来学习Koopman嵌入之外,Koopman理论也被应用于理解神经网络<ref>I. Manojlovic\', M. Fonoberova, R. Mohr, A. Andrejc\v uk, Z. Drmac\v , Y. Kevrekidis, and I. Mezic\', Applications of Koopman Mode Analysis to Neural Networks, preprint, https://arxiv.org/abs/2006.11765, 2020. (Cited on p. 290)</ref><ref>A. S. Dogra and W. T. Redman, Optimizing neural networks via Koopman operator theory, in Advances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, eds., Curran Associates, 2020, pp. 2087--2097. (Cited on p. 290)</ref>以及更广泛的算法<ref>F. Dietrich, T. N. Thiem, and I. G. Kevrekidis, On the Koopman operator of algorithms, SIAM J. Appl. Dyn. Syst., 19 (2020), pp. 860--885, https://doi.org/10.1137/19M1277059. (Cited on p. 290)</ref>。 | | 尽管Koopman数据驱动算法前景广阔,但要选取合适的观测函数仍然是一个核心挑战。即使对于相对简单的动力系统,Koopman 算子的特征函数也可能任意复杂,并且在有限基中只能被近似地表示。深度学习非常适合表示这种任意复杂的函数,并且最近在Koopman 预测方面显示出了巨大的潜力<ref>C. Wehmeyer and F. Noe\', Time-lagged autoencoders: Deep learning of slow collective variables for molecular kinetics, J. Chem. Phys., 148 (2018), pp. 1--9. (Cited on pp. 284, 290, 291)</ref><ref>N. Takeishi, Y. Kawahara, and T. Yairi, Learning Koopman invariant subspaces for dynamic mode decomposition, in Advances in Neural Information Processing Systems, 2017, pp. 1130--1140. (Cited on pp. 284, 290)</ref>。除了利用神经网络来学习Koopman嵌入之外,Koopman理论也被应用于理解神经网络<ref>I. Manojlovic\', M. Fonoberova, R. Mohr, A. Andrejc\v uk, Z. Drmac\v , Y. Kevrekidis, and I. Mezic\', Applications of Koopman Mode Analysis to Neural Networks, preprint, https://arxiv.org/abs/2006.11765, 2020. (Cited on p. 290)</ref><ref>A. S. Dogra and W. T. Redman, Optimizing neural networks via Koopman operator theory, in Advances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, eds., Curran Associates, 2020, pp. 2087--2097. (Cited on p. 290)</ref>以及更广泛的算法<ref>F. Dietrich, T. N. Thiem, and I. G. Kevrekidis, On the Koopman operator of algorithms, SIAM J. Appl. Dyn. Syst., 19 (2020), pp. 860--885, https://doi.org/10.1137/19M1277059. (Cited on p. 290)</ref>。 |
| − | 同样,如果矩阵 <math>\mathbf{K}</math> 被对角化,那么嵌入函数 <math>\varphi</math> 就对应于 Koopman 特征函数<ref>255, 147</ref>。变分自动编码器也被用于随机动力系统,例如分子动力学,其中从潜在变量映射回物理构型空间的过程是概率性的<ref>444, 274</ref>。相比之下,第二种范式,即测量值被提升到更高维空间,与机器学习中的许多结果相关,例如 Cover 定理<ref>232</ref>,即非线性问题在更高维嵌入中往往变得更加线性。 | + | 同样,如果矩阵 <math>\mathbf{K}</math> 被对角化,那么嵌入函数 <math>\varphi</math> 就对应于 Koopman 特征函数<ref name=lusch_2018>B. Lusch, J. N. Kutz, and S. L. Brunton, Deep learning for universal linear embeddings of nonlinear dynamics, Nature Commun., 9 (2018), art. 4950. (Cited on pp. 235, 284, 290, 291, 307)</ref>。[[变分自编码器]]也被用于随机动力系统,例如分子动力学,其中从潜在变量映射回物理构型空间的过程是概率性的。相比之下,第二种范式,即测量值被提升到更高维空间,与机器学习中的许多结果相关,例如[[Cover定理]]<ref>H. Lange, S. L. Brunton, and J. N. Kutz, From Fourier to Koopman: Spectral methods for long-term time series prediction, J. Mach. Learn. Res., 22 (2021), pp. 1--38. (Cited on pp. 235, 290, 291)</ref>,即非线性问题在更高维嵌入中往往变得更加线性。 |