In this paper, a new gene learning algorithm for optimum search problem is proposed, which extended the binary population-based incremental learning (PBIL) and selfish algorithm (SA) by allowing a gene's allele to be multi-valued. In this new algorithm, the entropy of probability distribution as used as the criterion of termination, and the evolution process is combined with local heuristic search. Three typical combinatorial optimization problems (maximum cut problem, scheduling problem and travelling salesman problem) are solved and some results are better than the best result of existing algorithm.