Learning of a Bayesian Autonomous Driver Mixture-of-Behaviors (BAD MoB) Model
Mark Eilers, Claus Moebus · Advances in human factors and ergonomics series · 2010
The Human or Cognitive Centered Design (HCD) of intelligent transport systems requires computational Models of Human Behavior and Cognition (MHBC). They are developed and used as driver models in traffic scenario simulations and risk- based design. The conventional approach is first to develop handcrafted control-theoretic or artificial intelligence based prototypes and then to evaluate ex post their learnability, usability, and human likeness. We propose a machine-learning alternative: The Bayesian estimation of MHBCs from behavior traces. The learnt Bayesian Autonomous Driver (BAD) models are empirical valid by construction. An ex post evaluation of BAD models is not necessary. BAD models can be built so that they decompose or compose skills into or from basic skills: BAD Mixture-of-Behaviors (BAD MoB) models. We present an efficient implementation which is able to control a simulated vehicle in real-time. It is able to generate complex behaviors of several layers of expertise by mixing and sequencing simpler behavior models.