Unsupervised Equalization over Adaptive Networks
Adaptive Networks comprise a group of agents capable of collecting and processing data
and sharing information across the network to collaboratively solve a task. This work
proposes implementing the Single Channel ICA algorithm within the adaptive networks
framework, referred to as AN-SCICA, for distributed filter estimation in blind equalization.
This algorithm is a technique from a well-studied area of signal processing known as
Blind Source Separation, and it has the advantage of being unsupervised while making
only weak assumptions about the data, which makes it well-suited to broader scenarios
of source distributions. Two implementations of the AN-SCICA are proposed: one for a
single-task scenario, where the channel is unique across the network and, therefore, the
agents estimate the same equalization filter, and another for a multitask scenario, where
the agents estimate a different yet related set of parameters. For the latter, the channels
will be treated as correlated across agents, and to perform equalization, this work proposes
a two-cascade distributed filter in which the agents share the parameters of the first stage
and individually estimate the second stage. The simulation results show the advantage
of the AN-SCICA over the standalone implementation of the algorithm through a broad
set of synchronous and asynchronous scenarios. Additionally, this work proposes a simple
strategy for reducing data exchange while maintaining acceptable network performance
for the AN-SCICA and the two-filter cascade for the multitask Adaptive Networks using
the distributed Constant Modulus Algorithm for unsupervised equalization