Modern video games come with highly realistic graphics enabling the players to interact with visually rich virtual worlds. Realistic (life-like) animation of nonplayer characters (NPCs) in such virtual worlds is particularly important to enhance the gaming experience. Multiagent systems are one effective approach to synthesize life-like behaviors and interactions by codifying simple rules into NPCs (each NPC as an autonomous agent). However, such behaviors generally come at the cost of increasing computational expense and complexity in terms of aspects such as number of rules and parameters. Therefore, the desire for high fidelity (highly realistic) behaviors is often in conflict with the drive for low complexity. Multiobjective evolutionary algorithms provide a sophisticated mechanism to optimize two or more conflicting objectives simultaneously. However, evolutionary computing techniques need an appropriate objective function to drive the exploration in the correct direction. Pairing of evolutionary techniques and multiagent systems is challenging in the classes of problems in which the fitness is evaluated based on human aesthetic judgment rather than on objective forms of measurements. In this study, we present a multiobjective evolutionary framework to evolve low complexity and high fidelity multiagent systems by utilizing a machine learning system trained by bootstrapping human aesthetic judgment. We have gathered empirical data in three problem areas-simulation of conversational group dynamics, sheepdog herding behaviors, and traffic dynamics, and show the effectiveness of our approach in deriving low complexity and high fidelity multiagent systems. Further, we have identified common properties of the Pareto-optimal frontiers in the three problem areas that can ultimately lead to an understanding of a relationship between simulation model complexity and behavior fidelity. This understanding will be useful in deciding which level of behavioral fidelity is required for the characters in video games based on the distance to the camera, importance to the scene, and available computational resources.