Calibration of stochastic radio channel models using kernels

3 Feb 2021

Rough Path Interest Group

Abstract

Radio channel modeling aims at replicating the behavior of the environment in which a radio signal propagates. Stochastic models of the radio channel are necessary simulation tools for designing and testing communication systems. These stochastic models simulate time-series data that is driven by an underlying point process whose points are not observable, thus rendering the likelihood function intractable. Estimating the parameters of the underlying point process is therefore a challenging task. We attempt to tackle this problem using approximate Bayesian computation (ABC) which is a likelihood-free inference framework. ABC relies on comparing summary statistics of the simulated data and the observed data in some distance metric. Parameter values that yield simulated data “close” to the observed data form a sample from the approximate posterior distribution. We make use of the maximum mean discrepancy, which is a notion of distance between probability distributions, as the distance metric in ABC. The proposed method is able to accurately estimate the parameters of stochastic channel models in simulation as well as when applied to real measurements.