Adaptive near minimum error rate training for neural networks with application to multiuser detection in CDMA communication systems

Adaptive training of neural networks is typically done using some stochastic gradient algorithm that aims to minimize the mean square error (MSE). For many classification applications, such as channel equalization and code-division multiple-access (CDMA) multiuser detection, the goal is to minimize the error probability. For these applications, adopting the MSE criterion may lead to a poor performance. A nonlinear adaptive near minimum error rate algorithm called the nonlinear least bit error rate (NLBER) is developed for training neural networks for these kinds of applications. The proposed method is applied to downlink multiuser detection in CDMA communication systems. Simulation results show that the NLBER algorithm has a good convergence speed and a small-size radial basis function network trained by this adaptive algorithm can closely match the performance of the optimal Bayesian multiuser detector. The results also confirm that training the neural network multiuser detector using the least mean square algorithm, although generally converging well in the MSE, can produce a poor error rate performance.