Neural Importance Sampling

Müller, Thomas; McWilliams, Brian; Rousselle, Fabrice; Groß, Markus; Novák, Jan

doi:10.1145/3341156

Cited by 229 publications

(202 citation statements)

References 30 publications

Supporting

Mentioning

202

Contrasting

Order By: Relevance

“…A very powerful definition of C was introduced in Ref. [25]. Both the domain and codomain of each Coupling Layer are defined to be the unit hypercube.…”

Section: B Piecewise Polynomial and Rational Quadratic Spline Couplimentioning

confidence: 99%

“…In particular, Ref. [25] experiments with piecewise linear and piecewise quadratic coupling transforms. In the implementation of a piecewise quadratic coupling transform, the bin width is allowed to vary in order to increase the flexibility of the Coupling Layer.…”

Section: B Piecewise Polynomial and Rational Quadratic Spline Couplimentioning

confidence: 99%

“…In this study we focus on Refs. [25,26] which generalizes the design of the Coupling Layers proposed by Ref. [28].…”

mentioning

confidence: 99%

See 2 more Smart Citations

Event generation with normalizing flows

et al. 2020

View full text Add to dashboard Cite

We present a novel integrator based on normalizing flows which can be used to improve the unweighting efficiency of Monte-Carlo event generators for collider physics simulations. In contrast to machine learning approaches based on surrogate models, our method generates the correct result even if the underlying neural networks are not optimally trained. We exemplify the new strategy using the example of Drell-Yan type processes at the LHC, both at leading and partially at nextto-leading order QCD. I. INTRODUCTIONNumerical simulation programs are a cornerstone of collider physics. They are used for the planning of future experiments, analysis of current measurements and, finally, reinterpretation based on an improved theoretical understanding of nature. They employ Monte Carlo methods to link theory and experiment by generating virtual collider events, which can then be analyzed like actual events observed in detectors [1,2].With more and more data available from the Large Hadron Collider (LHC) and the high-luminosity upgrade, the task of simulating collisions at high precision becomes a matter of concern for the high-energy physics community. The projected amount of computational resources falls far short of the needs for precision event generation [3]. Past studies of the scaling behavior of multi-jet simulations have shown that the compute needs are largely determined by the gradually decreasing unweighting efficiency [4,5]. Except for dedicated integrators, which require a detailed understanding of the physics problem at hand, adaptive Monte-Carlo methods seem the only choice to address this problem [6][7][8][9][10][11][12][13].With the rise of machine learning, this topic has seen a resurgence of interest recently. The possibility of using these techniques for integration in high-energy physics was first discussed in Ref. [14]. Boosted Decision Trees and Generative Adversarial Networks (GANs) were investigated as possible general purpose integrators. This new technique improved the integration of non-separable high dimensional functions, for which traditional algorithms failed. The first true physics application was presented in Ref. [15]. The authors used Dense Neural Networks (DNN) in order to perform a variable transformation and demonstrate that they obtain significantly larger efficiencies for three body decay integrals than standard approaches [16]. The major drawback of this method is its computational cost. Since the network acts as a variable transformation, its gradient must be computed for each inference point in order to determine the Jacobian. This becomes computationally heavy for high multiplicity processes.A completely orthogonal approach utilizes machine learning techniques directly for amplitude evaluation [17] or event generation [18][19][20][21][22][23][24]. Training data for these approaches are obtained from traditional event generation techniques, and hence the problem of efficient event generation still remains. In addition, one needs to ensure that the neural networks are trained well ...

show abstract

“…A very powerful definition of C was introduced in Ref. [25]. Both the domain and codomain of each Coupling Layer are defined to be the unit hypercube.…”

Section: B Piecewise Polynomial and Rational Quadratic Spline Couplimentioning

confidence: 99%

Section: B Piecewise Polynomial and Rational Quadratic Spline Couplimentioning

confidence: 99%

See 1 more Smart Citation

Event generation with normalizing flows

et al. 2020

View full text Add to dashboard Cite

show abstract

“…In order to be usable for multi-channel sampling, our adaptive model needs to be invertible. For this reason, we adopt the "Neural Importance Sampling" algorithm of [33]. The method of using a trainable mapping to redistribute the random numbers going into the generation of a sample is similar to how VEGAS is often used in practice.…”

Section: Neural-network Assisted Importance Samplingmentioning

confidence: 99%

“…These are adjusted by training neural networks, which has originally been proposed in [32]. Our work is in principle an application of 'Neural Importance Sampling' [33] as we employ the 'polynomial coupling layers' introduced therein, although we want to point out that the usage of coupling layers for importance sampling has also been studied in [34]. In Sec.…”

Section: Introductionmentioning

confidence: 99%

Exploring phase space with Neural Importance Sampling

et al. 2020

View full text Add to dashboard Cite

We present a novel approach for the integration of scattering cross sections and the generation of partonic event samples in high-energy physics. We propose an importance sampling technique capable of overcoming typical deficiencies of existing approaches by incorporating neural networks. The method guarantees full phase space coverage and the exact reproduction of the desired target distribution, in our case given by the squared transition matrix element. We study the performance of the algorithm for a few representative examples, including top-quark pair production and gluon scattering into three-and four-gluon final states.

show abstract