An　effective　and　fast　minimization　approach　is　proposed　for　the　prediction　of　protein　folding,　in　which　the　＇relative　entropy＇　is　used　as　a　minimization　function　and　the　off-　lattice　model　is　used.　In　this　approach,　we　only　use　the　information　of　distances　between　the　consecutive　C-alpha　atoms　along　the　peptide　chain　and　a　generalized　form　of　the　contact　potential　for　20　types　of　amino　acids.　Tests　of　the　algorithm　are　performed　on　the　real　proteins.　The　root　mean　square　deviations　of　the　structures　of　eight　folded　target　proteins　versus　the　native　structures　are　in　a　reasonable　range.　In　principle,　this　method　is　an　improvement　on　the　energy　minimization　approach.