In this paper, a new kind of swarm-based metaheuristic search method, called Elephant Herding Optimization (EHO), is proposed for solving global optimization tasks. EHO is inspired by the herding behavior of elephant group. These behaviors can be modelled into two following operators: clan updating operator and separating operator. In EHO, the elephants in each clan are updated according to its current position and matriarch through clan updating operator. It is followed by the implementation of the separating operator which can significantly enhance the population diversity at the later run phase of the search. EHO has been benchmarked by fifteen basic test problems, five IEEE CEC 2005 test cases, and two CEC 2011 problems in comparison with BBO, DE and GA. The results show that EHO is able to find the better function values on most benchmark problems than three algorithms.