Validation of agent based crowd egress simulation

Crowd behavior simulation has been an active field of research because of its utility in several applications such as emergency planning and evacuations, designing and planning pedestrian areas, subway or rail-road stations, besides in education, training and entertainment. The most advanced and realistic simulation systems employ intelligent autonomous agents with a balance between individual and group intelligence for scalability of the architectures. Although several systems have even been commercialized, little attention has been accorded to the problem of validating the outcomes of these simulations in a generalized manner, against reality. The extent of validation fails to go much beyond visual matching between the simulation and the actual scenarios (with recordings of human crowds), which can lead to highly subjective and often questionable conclusions. The existing numerical measures often rely on ad-hoc applications, e.g., local crowd densities are measured to verify patterns, without a systematic procedure to identify at what times in the simulation and the scenarios can the densities be compared. Furthermore, if there are multiple systems that simulate crowd behavior in the same scenario in the same virtual environment, then no technique is currently known to quantitatively compare these systems in terms of realism. In this abstract, we present the first (to the best of our knowledge) principled, unified, and automated technique to quantitatively validate and compare the global performance of crowd egress simulation systems. We also evaluate a multiagent based crowd egress simulation system (that we have recently developed, but we do not discuss this system here) using our technique and demonstrate a high degree of validity of that system as well.