A model of human psychology could help self-driving cars interact with human drivers on the road, according to a study. Gustav Markkula and colleagues combined several computational psychological models into one master-model to simulate pedestrians attempting to cross a busy road and the human drivers on that road. The goal of the model was to capture the underlying cognitive mechanisms responsible for observed behavior. Computational models of Bayesian perception, theory of mind, behavioral game theory, long-term valuation of action alternatives, and evidence accumulation decision-making were integrated to predict subtle behaviors observed in real traffic, such as hesitating before crossing in front of a yielding car, stopping a car in an exaggerated manner to signal to a pedestrian that the car will yield, or speeding up a car to deny a pedestrian a crossing opportunity. The resulting model also accurately predicted crossing behavior from 32 pairs of participants interacting as driver and pedestrian in a driver-pedestrian simulator. However, the authors note that this crossing situation scarcely scratches the surface of real-world driving. The complexity required to accurately model the myriad interactions between human beings in traffic will require significant further work. Machine learning may help model behavior that is beyond the current limits of mechanistic modeling, and combining the two approaches is particularly promising, according to the authors.
PNAS Nexus
Explaining human interactions on the road by large-scale integration of computational psychological theory
20-Jun-2023