Mixture of Behaviors in a Bayesian Autonomous Driver Model

The Human Centered Design (HCD) of Partial Autonomous Driver Assistance Systems (PADAS) requires Digital Human Models (DHMs) of human control strategies for traffic scenario simulations. We present a probabilistic model architecture for generating descriptive models of human driver behavior: Bayesian Autonomous Driver (BAD) models. They implement the sensory-motor system of human drivers in a psychological motivated mixture-of-experts (= mixture-of-schema) architecture with autonomous and goal-based attention allocation processes. Under the assumption of stationary behavioral processes models are specified across at least two time slices. Learning data are time series of relevant variables: percepts, goals, and actions. We can represent individual or groups of human and artificial agents. Models propagate information in various directions. When working top-down, goals emitted by a cognitive layer select a corresponding expert (schema), which propagates actions, relevance of areas of interest (AoIs) and perceptions. When working bottom-up, percepts trigger AoIs, actions, experts and goals. When the task or goal is defined and the model has certain percepts evidence can be propagated simultaneously top-down and bottom-up and the appropriate expert (schema) and its behavior can be activated. Thus, the model can be easily extended to implement a modified version of the SEEV visual scanning or attention allocation model of Horrey, Wickens, and Consalus. In contrast to Horrey et al. the model can predict the probability of attending a certain AoI on the basis of single, mixed, and even incomplete evidence (goal priorities, percepts, effort to switch between AoIs). In this paper we present the architecture and a proof of concept with plausible but artificial data.