Previous implementations of Genetic Algorithms in direct-space strategies for structure solution from powder diffraction data have employed the operations of mating, mutation and natural selection, with the fitness of each structure based on comparison between calculated and experimental powder diffraction patterns (we define fitness as a function of weighted-profile R-factor R-wp). We report an extension to this method, in which each structure generated in the Genetic Algorithm is subjected to local minimization of R-wp with respect to structural variables. This approach represents an implementation of Lamarckian concepts of evolution, and is found to give significant improvements in efficiency and reliability.