Learning Parameters of Stochastic Radio Channel Models From Summaries

Abstract

Estimating parameters of stochastic radio channel models based on new measurement data is an arduous task usually involving multiple steps such as multipath extraction and clustering. We propose two different machine learning methods, one based on approximate Bayesian computation (ABC) and the other on deep learning, for fitting stochastic channel models to data directly. The proposed methods make use of easy-to-compute summary statistics of measured data instead of relying on extracted multipath components. Moreover, the need for post-processing of the extracted multipath components is omitted. Taking the polarimetric propagation graph model as an example stochastic model, we present relevant summaries and evaluate the performance of the proposed methods on simulated and measured data. We find that the methods are able to learn the parameters of the model accurately in simulations. Applying the methods on 60 GHz indoor measurement data yields parameter estimates that generate averaged power delay profile from the model that fits the data.

A General Method for Calibrating Stochastic Radio Channel Models With Kernels

“…I for approximate estimates of the parameters after T = 10 iterations. The estimates are very similar to the ones reported in [22] where the polarized PG model was calibrated on data from the same measurement campaign. The estimate of P vis is almost one, indicating that nearly all scatterers are connected.…”

“…Considering that the log of the temporal moments are well modeled by a Gaussian distribution, we take s obs to be the vector consisting of the sample means and sample covariances of the elements of z, similar to [22]. In total, s obs consists of (I 2 + 3I)/2 elements for I temporal moments.…”

“…As a means to explore the posterior distribution over the parameter space efficiently, we employ a sequential Monte Carlo technique called PMC [20], [22], [52]. In PMC, the current parameter valuesΘ are used to generate a new set of parameters for the next iteration of the algorithm through importance sampling.…”

“…We encounter this to be a non-trivial task, and it may not even be possible for the more elaborate channel models. Similar problems exist in [20]- [22] where a low-dimensional vector of statistics should be redesigned or trained using an autoencoder [23] for the channel model at hand, which is not always trivial and may not generalize to other models. Moreover, summarizing the data leads to information loss that can hamper the accuracy of the parameter estimates.…”

Maximum Likelihood Calibration of Stochastic Multipath Radio Channel Models

“…Another potential weakness of such twostep procedures is that the resulting parameter estimates are highly sensitive to the estimation accuracy of the particular set of extracted multipath components. Calibration techniques that do not require multipath extraction but rely on summarizing the data into a set of statistics have been introduced recently in the literature [19]- [23]. However, these methods call for definition of appropriate summary statistics that are informative regarding the model parameters.…”