Comparison of hill-climbing and genetic algorithms for digital predistortion models sizing

Generalized Memory Polynomial (GMP) models are widely used for the linearization of power amplifiers. They offer a good tradeoff between linearization performance and implementation complexity. Their structure is defined by 8 integer parameters representing different non-linearity orders and memory lengths. These 8 degrees of freedom allow achieving very good linearization performance with a small number of coefficients. But the optimal sizing (determination of the 8 parameters) of such models could require huge computation, for instance, if these 8 parameters are bounded between 1 and 10, there are 10
⁸
models to test using an exhaustive search, which is very computationally heavy and time consuming. Therefore optimization algorithms are needed to search for a GMP model structure which provides a good tradeoff between modeling accuracy and complexity. In this paper, we compare two heuristic optimization algorithms, hill-climbing and integer genetic algorithms, in terms of convergence speed, and optimality of the obtained solution regarding the defined criterion. They are evaluated using data measurements from an LDMOS Doherty Power Amplifier dedicated to base stations. The results show that both algorithms allow decreasing very significantly the searching time while giving optimal or close to optimal solutions. Compared with hill-climbing, the genetic approach leads to a more difficult control and interpretation of the path followed by the search algorithm since it is based on random operations (crossovers and mutations).