AquaSense: Automated Sensitivity Analysis of Probabilistic Programs via Quantized Inference
AquaSense: Automated Sensitivity Analysis of Probabilistic Programs via Quantized Inference
复制标题
DOI：
10.1007/978-3-031-45332-8_16
复制
发表时间：
2023
影响因子：
--
通讯作者：
Zitong Zhou;Zixin Huang;Sasa Misailovic
中科院分区：
文献类型：
--
作者：
Zitong Zhou;Zixin Huang;Sasa Misailovic
我们提出了一种新颖的工具 AquaSense，用于自动推理概率程序的敏感性分析。在概率程序的背景下，敏感性分析研究先验分布参数的扰动如何影响程序的结果，即程序的后验分布。 AquaSense 利用量化推理，这是一种高效且准确的近似推理算法，可表示具有量化间隔的随机变量的分布。 AquaSense 是第一个支持概率程序敏感性分析的工具，同时具有符号性、可微分性和实用性。我们的评估将 AquaSense 与现有系统 PSense（完全依赖于符号推理的系统）进行了比较。 AquaSense 可以计算 12 个程序的所有 45 个参数的灵敏度，而 PSense 计算的灵敏度为 11/45。 AquaSense 对于连续分布的程序特别有效：它比 PSense 平均加速 18.10（相比之下，PSense 只能解决少数问题）。我们的评估表明 AquaSense 在离散程序上计算出准确的结果。对于 91% 的评估连续参数，AquaSense 在 40 秒内以高精度计算出灵敏度结果（误差低于 5%）。该论文还讨论了 AquaSense 的性能与准确性权衡，它可以为具有不同输入数据大小的程序提供不同的操作点。
We propose a novel tool, AquaSense, to automatically reason about the sensitivity analysis of probabilistic programs. In the context of probabilistic programs, sensitivity analysis investigates how the perturbation in the parameters of prior distributions affects the program’s result, i.e., the program’s posterior distribution. AquaSense leverages quantized inference, an efficient and accurate approximate inference algorithm that represents distributions of random variables with quantized intervals. AquaSense is the first tool to support sensitivity analysis of probabilistic programs that is at the same time symbolic, differentiable, and practical.Our evaluation compares AquaSense with an existing system PSense (a system that relies on fully symbolic inference). AquaSense can compute the sensitivity of all 45 parameters from 12 programs, compared to 11/45 that PSense computes. AquaSense is particularly effective on programs with continuous distributions: it achieves an average speedup of 18.10over PSense (which, in contrast, can solve only a handful of problems). Our evaluation shows that AquaSense computes exact results on discrete programs. On 91% of evaluated continuous parameters, AquaSense computed the sensitivity results within 40 s with high accuracy (below 5% error). The paper also discusses AquaSense’s performance-accuracy trade-offs, which can enable different operational points for programs with different input data sizes.