Dynobench - extended Strogatz benchmark for system identification methods

The dynobench repository contains a benchmark for system identification methods. Currently includes models of 10 dynamical systems: Bacterial respiration, Bar magnets, Glider, Lotka-Volterra, Predator-Prey, Shearflow and Van der Pol from the Strogatz dataset, as well as Lorenz, Coupled phase oscillators and Stuart-Landau models for dynamical systems that often appear in the research community. They also add variety to the benchmark as the Lorenz oscillator model introduces a larger set of state variables (three compared to two), and the coupled phase oscillators model is non-autonomous, which is reflected in the explicit incorporation of time in its equations. The repository contains the 'data' folder, where the simulations of ten dynamical systems are stored, simulated under 6 different configurations of data quality. The first dimention modifies the data length and coarseness, where a 'small' dataset includes simulations of 10 seconds with a 0.1 sampling step, and a 'large' dataset includes simulations of 20 seconds with a 0.01 sampling step. The second dimention of data quality modifies the amount of noise in the data, where there are three levels of noise (no noise, moderate levels with 30 dB signal-to-noise ratio and high levels of noise with 13 dB signal-to-noise ratio). The data can be used by itself, without the need to look at the python code. The repository also contains the main.py script by which the data can be generated. The 'src' folder contains additional python scripts that are needed to generate the data. The data were created by first randomly setting the initial values for one category, in particular a configuration of 'small', 'noise-free' and 'train' data (using inits_type = "random"). Then, all the other configurations were generated by using the same initial values. Inside the script main.py there is more information about the settings and how to run the script. The benchmark was created as a part of the research described in the paper titled Probabilistic grammars for modeling dynamical systems from coarse, noisy, and partial data, written by Omejc et al. (in submission).dynobench存储库包含一套面向系统辨识方法的基准测试集。目前已收录10种动力学系统的模型：取自Strogatz数据集的细菌呼吸动力学模型、条形磁铁动力学模型、滑翔机动力学模型、洛特卡-沃尔泰拉模型、捕食者-猎物模型、剪切流模型以及范德波尔振子模型，此外还有科研领域常见的洛伦兹振子、耦合相位振子与斯图尔特-兰道模型。上述模型为基准测试增添了丰富性：洛伦兹振子模型拥有更多的状态变量（3个，相较于多数模型的2个），而耦合相位振子模型为非自治系统，其方程中显式引入了时间项，这一特性清晰体现了这一点。 该存储库包含`data`文件夹，其中存储了10种动力学系统的仿真数据，这些数据基于6种不同的数据质量配置生成。第一个维度用于调整数据长度与采样粗粒度：「小」数据集包含时长10秒、采样步长0.1的仿真数据，「大」数据集则包含时长20秒、采样步长0.01的仿真数据。第二个数据质量维度用于调整数据中的噪声水平，共包含三级噪声设置：无噪声、中等噪声（信噪比30 dB）以及高噪声（信噪比13 dB）。该数据可独立使用，无需查阅Python代码。 该存储库还包含用于生成数据的`main.py`脚本；`src`文件夹中包含生成数据所需的其他Python脚本。数据生成流程为：首先为一组基准配置（即「小尺寸、无噪声、训练集」数据，采用`inits_type = "random"`参数）随机设置初始值，随后所有其他配置均沿用该初始值完成生成。`main.py`脚本中包含更多关于参数设置与脚本运行方式的详细说明。 本基准测试是Omejc等人投稿待发表的论文《基于粗粒度、含噪声与不完整数据的动力学系统建模概率语法》（Probabilistic grammars for modeling dynamical systems from coarse, noisy, and partial data）中所述研究的组成部分。