Abstract:We present Neural Quantile Estimation (NQE), a novel Simulation-Based Inference (SBI) method based on conditional quantile regression. NQE autoregressively learns individual one dimensional quantiles for each posterior dimension, conditioned on the data and previous posterior dimensions. Posterior samples are obtained by interpolating the predicted quantiles using monotonic cubic Hermite spline, with specific treatment for the tail behavior and multi-modal distributions. We introduce an alternative definition for the Bayesian credible region using the local Cumulative Density Function (CDF), offering substantially faster evaluation than the traditional Highest Posterior Density Region (HPDR). In case of limited simulation budget and/or known model misspecification, a post-processing calibration step can be integrated into NQE to ensure the unbiasedness of the posterior estimation with negligible additional computational cost. We demonstrate that NQE achieves state-of-the-art performance on a variety of benchmark problems.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to perform Simulation - Based Inference (SBI) efficiently and accurately in the case of a limited simulation budget and/or when the known model is not completely correct. Specifically, the author proposes a new SBI method - Neural Quantile Estimation (NQE), which estimates the posterior distribution based on conditional quantile regression. NQE learns the one - dimensional quantiles of each posterior dimension in an autoregressive manner and uses the monotonically cubic Hermite spline - interpolated predicted quantiles to obtain posterior samples. In addition, the paper also proposes a new method for defining Bayesian credible regions based on the local Cumulative Distribution Function (CDF), as well as a post - processing calibration strategy to ensure the unbiasedness of posterior estimates. ### Main contributions of the paper 1. **New SBI method**: Introduced NQE, an SBI method based on conditional quantile regression, which can provide efficient posterior estimates under a limited simulation budget. 2. **Posterior sample generation**: Generate posterior samples by interpolating the predicted quantiles, specifically handling tail behavior and multimodal distributions. 3. **New definition of Bayesian credible regions**: Proposed a definition of Bayesian credible regions based on local CDF, which is faster to evaluate than the traditional Highest Posterior Density Region (HPDR). 4. **Post - processing calibration strategy**: In the case of a limited simulation budget and/or when the known model is not completely correct, ensure the unbiasedness of posterior estimates through a post - processing calibration step without additional computational cost. ### Method overview - **Quantile estimation and interpolation**: NQE estimates the quantiles of parameters given data by minimizing the weighted L1 loss. Then, these quantiles are interpolated using monotonically cubic Hermite splines to reconstruct the complete posterior distribution. - **Multidimensional parameter handling**: For multidimensional parameters, NQE applies the above method to each dimension in turn while considering the information of previous dimensions. - **Regularization**: To improve the smoothness of NQE predictions, a regularization scheme is introduced to improve the prediction results by penalizing the non - smooth parts in the PDF. - **Post - processing calibration**: Make the posterior distribution wider by adjusting the distance between quantiles, thereby ensuring the unbiasedness of posterior estimates. ### Experimental results The paper verifies the performance of NQE on multiple benchmark problems. The results show that NQE has reached the state - of - the - art level in various tasks and even outperforms existing methods in some cases. In particular, through post - processing calibration, NQE can ensure the unbiasedness of posterior estimates, even in the case of a limited simulation budget. ### Conclusion As a new SBI method, NQE not only performs well under a limited simulation budget but also ensures the unbiasedness of posterior estimates through the post - processing calibration strategy, providing a new solution for Simulation - Based Inference.

Simulation-Based Inference with Quantile Regression

Quantile Inference for Moderate Deviations from a Unit Root Model with Infinite Variance

Distributed Bootstrap Simultaneous Inference for High-Dimensional Quantile Regression

Cosmological Analysis with Calibrated Neural Quantile Estimation and Approximate Simulators

Quantile Regression Neural Networks: A Bayesian Approach

Censored Quantile Regression Neural Networks

Bayesian Non-parametric Quantile Process Regression and Estimation of Marginal Quantile Effects

A Stochastic Approximation Method for Simulation-Based Quantile Optimization

Bayesian nonparametric quantile process regression and estimation of marginal quantile effects

Online Quantile Regression

Learning Quantile Functions without Quantile Crossing for Distribution-free Time Series Forecasting

Semi-Supervised Quantile Estimation: Robust and Efficient Inference in High Dimensional Settings

A Quantile Neural Network Framework for Two-stage Stochastic Optimization

Bayesian Sequential Design of Computer Experiments for Quantile Set Inversion

Preconditioned Neural Posterior Estimation for Likelihood-free Inference

Neural Networks for Partially Linear Quantile Regression

Weighted Composite Quantile Inference for Nearly Nonstationary Autoregressive Models

Estimation of Non-Crossing Quantile Regression Process with Deep ReQU Neural Networks.

Low-Budget Simulation-Based Inference with Bayesian Neural Networks

Horseshoe Prior Bayesian Quantile Regression

Deep Learning for Quantile Regression under Right Censoring: DeepQuantreg