Blind channel and symbol estimation for wireless communications via an affinity neural network

We present a neural network (NN) approach to the blind channel and symbol estimation problem in portable communications. It is based on deterministic blind estimation methods, which utilize multiple antennas and/or oversampling in order to identify the channel and the data symbols. These deterministic approaches employ a least squares error metric, and then solve the problem algebraically. We use a NN to solve the estimation problem by mapping the quadratic cost function to the NN energy function, which is then minimized by iteratively updating each of its nodes (clusters of neurons called "affinity cells"). While its performance is found to be comparable to the SVD-based least squares methods, the NN offers significant practical advantages stemming from its distributed and fault-tolerant nature. Another important benefit of the NN approach, in contrast to the algebraic approaches, is its natural ability to yield the best solution in the space of finite word-length parameter vectors.