arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 1912.04412astro-ph.COastro-ph.GAastro-ph.IM
收藏

使用快速生成的训练数据的概率宇宙网络分类

Probabilistic cosmic web classification using fast-generated training data

Brandon Buncher, Matias Carrasco Kind

摘要:

我们提出了一种新的方法,强大的概率宇宙网络粒子分类在三维使用监督机器学习算法。使用简化的$Λ$ CDM玩具模型生成训练数据,该模型具有用于生成光晕、细丝和空隙的预定算法。虽然该框架不受物理建模的约束,但它可以比N体模拟更快地生成,而不会损失分类精度。对于该数据集中的每个粒子,测量局部密度场幅度和方向性。这些测量被用于训练随机森林算法,该算法被用于在$Λ$ CDM、具有$256^3$粒子的仅暗物质N体模拟以及另一玩具模型数据集中为每个粒子分配类别概率。通过使用不同的特征集比较每个数据集中分配给粒子的类别的ROC曲线和其他统计度量的趋势,我们证明了局部密度场幅度和方向性的测量的组合能够对晕,细丝,我们还表明,这种训练特征的组合确保了我们的玩具模型的构建不会影响分类。使用完全监督算法可以更好地控制被认为对分类重要的信息,防止深度学习模型中的超参数和模式崩溃引起的问题。由于训练数据生成的速度,我们的方法具有高度可扩展性,因此特别适合对大型数据集(包括观测数据)进行分类。

英文摘要:

We present a novel method of robust probabilistic cosmic web particle classification in three dimensions using a supervised machine learning algorithm. Training data was generated using a simplified $Λ$CDM toy model with pre-determined algorithms for generating halos, filaments, and voids. While this framework is not constrained by physical modeling, it can be generated substantially more quickly than an N-body simulation without loss in classification accuracy. For each particle in this dataset, measurements were taken of the local density field magnitude and directionality. These measurements were used to train a random forest algorithm, which was used to assign class probabilities to each particle in a $Λ$CDM, dark matter-only N-body simulation with $256^3$ particles, as well as on another toy model data set. By comparing the trends in the ROC curves and other statistical metrics of the classes assigned to particles in each dataset using different feature sets, we demonstrate that the combination of measurements of the local density field magnitude and directionality enables accurate and consistent classification of halo, filament, and void particles in varied environments We also show that this combination of training features ensures that the construction of our toy model does not affect classification. The use of a fully supervised algorithm allows greater control over the information deemed important for classification, preventing issues arising from hyperparameters and mode collapse in deep learning models. Due to the speed of training data generation, our method is highly scalable, making it particularly suited for classifying large datasets, including observed data.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑