Generating more Realistic Packet Loss Patterns for Wireless links using Neural Networks

Abstract

Simulations of wireless network connections are essential for
the development of new technologies because they are far
more scalable than real-world experiments and reproducible.
Modeling packet loss realistically provides a highly abstract
yet powerful tool for the simulation of wirelesses links. Typi-
cally, simple statistical models or replaying of recorded traces
are used for the simulation. For a proper parametrization of
simple statistical models, recorded traces are required, too.
Both approaches have drawbacks: replaying traces is limited
to the length of the traces, a repetition may lead to unwanted
effects in the simulation. The statistical models solve this, but
the resulting packet loss patterns significantly differ from real
ones. In this paper, we propose using a neural network in-
stead. It takes the same kind of input, i.e., a real-world trace,
but it can generate longer traces with more realistic loss pat-
terns. We share pre-trained neural networks for multiple links
in office and industry scenarios with the community for use
in future research.