Talk: Models and their Validation and their Role in Perception--and in Safe Autonomous Vehicles

John Rushby

Presented at a virtual meeting of IFIP WG 10.4 on 11 May 2022

Abstract

A model is a simplified description of something that can be used in explaining or predicting its behavior. I will begin by sketching some of the history of criteria and methods for validating models, their use in science and engineering, and the idea of model-based control.

I will then switch to animal perception, and the idea that it operates top-down, using models for prediction. This dates back to Helmholtz in the 1860s and was made explicit by Gregory in his 1980 paper "Perceptions as Hypotheses". It is now the dominant theory in many areas of CogSci, under the names "Predictive Processing" and "Free Energy".

I have previously argued that the perception components of autonomous systems should work in a similar way to maintain the models that are used to calculate the actions that control the system.

Here, I will briefly recap this and will then attempt to bring the threads together and suggest how we can assure these models and thereby deliver autonomous vehicles whose safety is at least conceivable.

Video

27 mins, mp4

Slides

PDF

Relevant Papers

Assessing Confidence in Assurance 2.0 by Robin Bloomfield and John Rushby, Tech Report SRI-CSL-2022-02, May 2022 and also available as arXiv 2205.04522

Model-Centered Assurance For Autonomous Systems by Susmit Jha, John Rushby, and N. Shankar. Presented at SafeComp, September 2020. Published in Springer LNCS 12234.

Having trouble reading our papers?
Return to John Rushby's bibliography page
Return to the Formal Methods Program home page
Return to the Computer Science Laboratory home page