Sampling the extrema from statistical models of music with variable neighbourhood search