16 Paper Presentations at ICDCN2026
16 presentations and 2 awards at ICDCN 2026 held in Nara, Japan (January 2026)
Publications & Talks Awards
International Conference · 2026
Neural Surrogate Model for Autonomous Driving Communications Based on Strategic Sampling
Abstract
Autonomous driving and vehicle-to-everything applications impose near-real-time requirements on communication-quality estimation, but high-fidelity simulators are computationally prohibitive for online use. To bridge this gap, neural surrogate models are used to emulate expensive simulators, enabling fast, deployable inference. However, training accurate surrogates itself demands many expensive simulation runs, often wasting compute on redundant simulations. In this paper, we present a methodological framework for constructing neural surrogate models that approximate communication performance under constrained simulation budgets. The proposed workflow combines global coverage via Latin Hypercube Sampling with gradient-based focused sampling to efficiently allocate expensive simulation runs to regions of greatest model uncertainty or sensitivity.
Research Note
Autonomous driving and vehicle-to-everything applications need to know the communication quality they can expect at a given place and moment, and they need to know it while the vehicle is moving. High-fidelity propagation simulators can answer that question accurately but take far too long to be queried online.
A surrogate model is the usual way around this. Instead of reproducing the physics, a neural network learns the mapping from input conditions to simulator output, so that inference afterwards costs almost nothing. The catch is that building the surrogate requires running the expensive simulator many times to generate training data, and sweeping conditions blindly tends to spend that budget on runs that add little information.
This work therefore treats the choice of which simulations to run as part of the design. The proposed workflow first uses Latin Hypercube Sampling to cover the input space evenly, then applies gradient-based focused sampling to concentrate the remaining budget on regions where the model is most uncertain or where the output is most sensitive to its inputs.
The contribution is a methodology for constructing communication surrogates under a constrained simulation budget rather than a single trained model.
Related Lab Activities
