Development of a paradigm to train and validate agent-based models

Agent-based modeling (ABM) is a computational technique employed to simulate the behavior of autonomous units called agents. An agent is a virtual entity which interacts with its environment and other agents and is capable of making decisions and taking actions based on these interactions. Relatively simple rules are utilized to govern individual interactions; however the cumulative behavior of all agents may be complex and may sometimes reveal interesting patterns. Agent-based modeling, though first introduced in the field of artificial intelligence, has gained acceptance in a many diverse fields such as finance engineering, bio-film modeling, social computation, and ecology. The systems studied in these fields are inherently complex involving many interacting components. These systems are modeled by representing the various system components as agents and having simple rules governing the behavior of agents. These rules can be changed to match the emergent behavior of the model with the behavior of the actual system. However, validating such a model is not straight forward as it is difficult to objectively determine similarity between the patterns in the model output and data. We propose a universal paradigm to facilitate such comparisons by extracting features from patterns in the data and the model output and use these features to train and validate the model. The merit of this methodology will be adopted to train an ABM simulating the movement of cattle grazing in a pasture using data from GPS units installed on individual cows.