Evolving Direct Perception Models of Human Behavior in Building Systems

Software agents that use direct (or active) perception of the environment have recently been shown to correspond well with pedestrian movement within building and urban systems. The algorithm, based on Gibson’s theory of affordances, combines random selection of destination from their field of view with reassessment of the destination every few steps. However, although the agents correlate with human movement on aggregate, as individuals they progress more erratically than people do. It might seem necessary to add higher cognitive functions in order to guide them more convincingly, but here we show that it is possible to improve their behavioral response through artificial evolution of their existing navigation rules. First we show that the destination selection method approximates stochastic direction choice by length of line-of-sight. Then we use the lines of sight to provide a set of inputs to the agents, or animats, which we evolve to fit human usage patterns within a building as best possible. We demonstrate that while agents using informational change inputs fail to evolve to fit movement patterns, an input that compares sight-line lengths improves models qualitatively, but not quantitatively, which further implies that the individual guidance mechanism may be independent of the instant spatial properties acting on direct perception.