Automatic mapping of human behavior data to personality model parameters for traffic simulations in virtual environments

Abstract

We present an approach towards automatic parameter identification for personality models in traffic simulation telemetry. To this end, we compare the behavior data of human and artificial drivers in the same virtual environment. We record the driving behaviors of human subjects in a car simulator and use evolutionary strategies to infer parameters of models of artificial drivers from the recorded data. We evaluate our approach in several prototypic traffic situations in which we compare the resulting artificial agents against human drivers as well as against simple baseline implementations of artificial drivers. As a result, we show that particular ranges of parameters of a driver profile can be inferred for which the simulated driving behavior does not change. We further show that precision depends on the amount of data and the scenarios in which these were recorded. The proposed method can also be applied to compare human and artificial driving behavior.