dysts
收藏资源简介:
dysts数据集是一个不断增长的混沌动力系统数据库,目前包含131个已知的混沌动力系统,这些系统跨越了天体物理学、气候学、生物化学等多个领域。每个系统都与预先计算的多变量和单变量时间序列配对。数据集的规模与现有的静态时间序列数据库相当,但我们的系统可以通过重新积分来生成任意长度和粒度的新数据集。数据集中的每个系统都注释有已知的数学属性,并进行了特征分析,以广泛分类集合中存在的不同动态。混沌系统本质上是挑战预测模型的,我们在广泛的基准测试中将预测性能与混沌程度相关联。此外,我们还利用数据集独特的生成特性进行了几个概念验证实验:代理转移学习以改善时间序列分类,重要性抽样以加速模型训练,以及基准符号回归算法。
The Dysts dataset is a growing database of chaotic dynamical systems. It currently contains 131 known chaotic dynamical systems spanning multiple disciplines including astrophysics, climatology, biochemistry, and other fields. Each system is paired with precomputed multivariate and univariate time series. The scale of this dataset is comparable to that of existing static time series databases, but it can generate new datasets of arbitrary length and granularity through numerical re-integration. Every system in the dataset is annotated with its known mathematical properties and subjected to feature analysis to broadly categorize the distinct dynamics present in the collection. Chaotic systems inherently pose challenges to predictive models, and we correlated prediction performance with the degree of chaos across extensive benchmark tests. Furthermore, we leveraged the unique generative capabilities of this dataset to conduct several proof-of-concept experiments: surrogate transfer learning to improve time series classification, importance sampling to accelerate model training, and benchmarking of symbolic regression algorithms.

- 1Chaos as an interpretable benchmark for forecasting and data-driven modelling德克萨斯大学奥斯汀分校物理与奥登研究所 · 2023年



