| This thesis discusses some techniques to improve the neural network based decision-directed blind equalization scheme. First, a Complex Back-Propagation algorithm is derived for Dual-Mode Modified Constant Modulus Algorithm; and to make it functional with MLP structure, a "gain-factor" is introduced to avoid local minima. A new dual-mode scheme is also proposed, which switches from blind to decision-directed mode without needing any gain-factor. Some new decision-directed blind equalization algorithms are also discussed. Second, an adaptive Stop-and-Go algorithm is proposed to give better tracking capability. Third, the idea of using two cascaded activation functions in the output neuron is proposed to obtain better correlation between the real and the imaginary parts of the output data and to get lower steady-state Mean-Square-Error. Fourth, the nonlinearity of the activation function is made adaptive based on the energy and kurtosis of the error, giving faster convergence speed and improved stability. Finally, a Recursive Least Square based Stop-and-Go Decision-Directed scheme is derived for the complex-valued multilayer perceptron structure. Proposed schemes, simulated on complex-valued channels for M-ary QAM signals, are showing excellent performance. |
