Optimal scaling of MaLa for non linear regression