Βι¶ΉΣ³» AI Seminar VI Nathaniel Virgo: On good regulators and their relationship to models
By: Aleks Kossowska
Last updated: Monday, 8 April 2024
Title: On good regulators and their relationship to models
Abstract:
A question of current interest is whether and in what sense AI systems develop 'world models'. A related question is whether and to what extent biological systems must model or represent their environment in order to carry out biological tasks - and if they do, what form such models take, and whether they can be said to be found "in the brain".
There are a number of classical results that say something along the lines of "if a system is an optimal solution to some particular kind of problem, then it must have something we can call a model." These include Conant and Ashby's celebrated 'good regulator theorem' and related results, as well as the so-called 'complete class theorems' in statistics. These results hint at a close relationship between optimality (or satisfactoriness) and model-having, but they are quite limited in scope, and if interpreted too broadly seem to admit counterexamples, such as Braitenberg vehicles and open-loop control.
I will present a result with a similar statement that is perhaps more suited to reasoning about embodied agents, and which can account for the apparent counterexamples. I will show how a notion of 'belief updating' arises that is analogous to Bayes' rule. I will also talk about progress towards understanding the relationship between optimality and models in a much more general sense.
Watch the recording of seminar given 13/03/2024: