Intelligent traffic signal control is necessary for improving traffic efficiency. These fast changing and challenging traffic scenarios and demands are generally handled by professional traffic engineers. However, it may take years of time and thousands of practices to train such an engineer. This paper proposes a deep behavioral cloning method to learn how to control the traffic signal effectively and efficiently from virtual expert demonstration. The method imitates promising working behavior of optimized offline solutions, and applies it to solve online traffic signal control problems of the similar scenario. Different traffic demand patterns are generated through a combination of different kinds of components. Then the virtual demonstration is constructed by getting an exclusive and optimized solution for each generated virtual traffic demand pattern through a heuristic random search method. After that, a deep neural network-based behavioral cloning method is employed to learn from the virtual demonstration and finish on-line traffic signal control task. The experimental results show that compared with other methods, the proposed method significantly reduces the waiting time and time loss in different situations. And the average traffic speed of the road network at different saturation levels can be improved by 1.58% to 11.54%.