Steering behavior of autonomous agents plays important roles in many simulation applications, such as simulation of pedestrian crowds, simulation of evacuation scenarios, simulation of ecosystems, simulation of autonomous robots, and simulation of artificial life in virtual environments used in computer games. It is desirable to have an approach that can automatically discover multiple candidate models for steering behavior simulation besides manual approach (trial-and-error fashion) and data-driven approach. Towards this goal, this work presents an approach that searches for candidate models of steering behavior in an automated way. The proposed framework includes two components. A model space specification provides a formal specification for a general structure from which various models can be constructed, and a search method to search for a set of candidate models based on requirements. To support more complex scenarios, we further add three major extensions including: (1) Activation component assign dynamic priorities for behaviors depending on surround environments. (2) Multiple search stages are provided to assist the evolutionary search algorithm to distribute computational resources better. (3) A special type of entity called space entity to assist agents receive information not only from other entities (agents, obstacles), but also from surrounding empty space. The approach is able to discover multiple candidate models for three basic steering behaviors including the leader- following ( Bleader_following), personal space maintenance ( Bpersonal_space), and mobile obstacle avoidance ( Bobstacle_avoidance). The results show that different possibilities of steering behavior support modelers to have a better understanding of the problem under study, hence assist modelers to develop more advanced models by testing different combinations of the basic steering behaviors. We evaluate all combinations between three basic steering behaviors including: (1) Bleader_following + Bobstacle_avoidance, (2) Bobstacle_avoidance + Bpersonal_space, (3) Bleader_following + Bpersonal_space,

and (4) Bleader_following + Bobstacle_avoidance + Bpersonal_space. We further test the approach with two variations of scenario 4: (5) The leader surrounding + Bpersonal_space, (6) Hall-way evacuation with an obstacle in the middle. The results show that the framework is also able to discover multiple models for each of these composite steering behaviors, and several of them have good scalability and robustness.