Agent-based models (ABMs) simulate agents and the interactions between agents and their environments. We focus on the coevolution of prey and predator ABM. This thesis proposes to train agent models for increased accuracy in scene modeling by selecting agents capable of recognizing and modeling the relative locations of landmarks in their simulated environment to test and demonstrate the feasibility of this method. There are, therefore, two goals to this thesis. First, to develop a simulation-based technique for training AI agents in proficiency of scene modeling using population level “survivability” metrics. Since scene modeling is viewed as a competitive advantage in terms of survival in nature, the research investigates whether a survival simulation is a sufficient motivation for selecting agents capable of accurately modeling their surroundings. Through experiments, we demonstrate the effectiveness or ineffectiveness of this ABM-inspired survival simulation as a viable training method for simulating complex behaviors. Second, we show how procedural generation techniques can extend existing ABMs into the third dimension and allow the rendering of this environment from several unique perspectives. This extension opens the door between simulated robotics and the large base of available ABM models, allowing new methods for comparing robot behaviors in highly dynamic environments.