Abstract: Heuristic hill-climbing search algorithm can do effectively pruning. In practice, it can be used to searching a large hypothesis space to get an optimal or approximate optimal solution. Beam search algorithm retains its advantage in efficiency while reducing the risk of converging to locally optimal hypotheses. Beam search algorithm is widespread used AI field. To k-size beam search, due to only k paths is maintained the key to optimize the accuracy of beam search is how to select the k paths. In most of search algorithms, the k candidates with the most high performance measure value are selected at each searching step. In this paper, the author presented some methods of candidate selection of beam search approaches, and the thought of avoiding “full of blood brother nodes” is presented. The experiments were done on the UCI repository of machine learning databases.