It is difficult to discover effective behavior for NPCs automatically. For instance, evolutionary methods can learn sophisticated behaviors based on a single objective, but realistic game playing requires different behaviors at different times. Such complex behavior is difficult to achieve. What is needed are multi-objective methods that reward different behaviors separately, and allow them to be combined to produce multi-modal behavior. While such methods exist, they have not yet been applied to generating multi-modal behavior for NPCs. This paper presents such an application: In a domain with noisy evaluations and contradictory fitness objectives, evolution based on a scalar fitness function is inferior to multi-objective optimization. The multi-objective approach produces agents that excel at the task and develop complex, interesting behaviors.