首页>
外文会议>AIAA aviation forum
>Probabilistic Aircraft Trajectory Prediction with Weather Uncertainties using Approximate Bayesian Variational Inference to Neural Networks
【24h】
Probabilistic Aircraft Trajectory Prediction with Weather Uncertainties using Approximate Bayesian Variational Inference to Neural Networks
A key consideration in Trajectory Prediction (TP) tools is the confidence that can be placed on the prediction. We purpose a non-deterministic TP neural network using tractable approximate Bayesian variational inference for the model parameters considering weather effects. This work adopts the state-of-art in Bayesian Deep Learning research and builds a neural network model with stochastic convolutional, recurrent, and fully-connect layers. The purposed stochastic variational method outperforms the dropout approximate to Variational Inference and performs reliable uncertainty estimates. It can be easily applied to most neural net architectures and also provides a simple pruning heuristic that can drastically reduce the number of model parameters compares to ensemble methods. The experiment is conducted with the Atlanta Air Route Traffic Control Center (ZTL) flight data and the corridor integrated weather system (CIWS) weather data from Sherlock Data Warehouse (SDW) on June 24th, 2019. The experimental results show better variance reduction than dropout-based methods. The uncertainty estimates are more reliable thanks to the Kullback-Leibler divergence (KL-divergence) term within the optimization objective.
展开▼
机译：轨迹预测（TP）工具中的关键考虑是可以放置在预测上的置信度。我们目的是一种非确定性TP神经网络，用于考虑天气效应的模型参数的近似贝叶斯变分推理。这项工作采用贝叶斯深度学习研究的最先进，并用随机卷积，复发和完全连接层构建神经网络模型。所用随机变分方法优于变分推理的差异差异，并且执行可靠的不确定性估计。它可以很容易地应用于大多数神经网络架构，并且还提供了一种简单的修剪启发式，可以大大减少模型参数的数量比较的融合方法。该实验与亚特兰大航线交通管制中心（ZTL）飞行数据和来自Sherlock Data Warehouse（SDW）的走廊综合天气系统（CIWS）天气数据进行了2019年6月24日。实验结果表明比辍学更好地减少基于基础的方法。由于优化目标内的Kullback-Leibler分歧（KL-Diversence）术语，不确定性估计更可靠。
展开▼
