Learning schemes for Genetic Programming

A learning capability is introduced in the Genetic Programming (GP) paradigm. This is achieved by enhancing GP with Simulated Annealing (SA), where the latter adapts the parameter values (in the form of node gains) in the structures evolved by the former. A special feature of this approach is that, due to the particularities of the representation used, it allows engineering problems (in which numerical parameters are important) to be addressed, thus extending the applicability of the GP paradigm. We study two different learning schemes, which we refer to as Darwinian and Lamarckian according to whether the learned node gains are inherited or not. We compare the results obtained by these two techniques to those obtained in the absence of learning (both with node gain representation and standard GP representation). The results show the great interest of both learning schemes. The application presented is a classical Digital Signal Processing problem: the equalisation of a noisy communications channel.