Скачать или смотреть Learning Active Inference Models of Perception and Control

Safety through Disruption (SAFE-D) UTC Webinar Series
Project: Learning Active Inference Models of Perception and Control
Presenter: Dr. Alfredo Garcia, Texas A&M Transportation Institute
Project: Safe-D 06-009

This project aimed to estimate models of drivers’ car-following behavior based upon active inference, a theoretical framework for perception and action that is grounded in cognitive science. We developed the models based upon a dataset consisting of the positions, velocities, and headings of all vehicles in a road segment at a given sampling frequency. These data are from a publicly available driving dataset recorded using drones on fixed road segments in the USA, Germany, and China. Two types of models were compared and the results showed that an active inference-based model outperformed an imitation learning-based models from the machine learning literature. The results were used to perform counterfactuals (i.e., predict behavior with observation-action sequences not in the dataset) and trace those behaviors back to perception and driver preferences. The results indicate that learning active-inference models from data is a promising alternative to black-box models of driving.

More information: https://safed.vtti.vt.edu/projects/en...