NME130/Uncertainty
The following summarizes opinions and notes of Andy, Nader, and myself.
The main two topics of the week were (i) uncertainty and (ii) Bayesian methods and graphical models. Although these seem related, we were not able to bring a unified story about these two topics (probably because our lack of knowledge in (ii)). Therefore, we tried to give a two-folded message:
(1) Uncertainty as it appears in controls, optimization, computer science, and networking is too general to cover as a separate topic and multiple aspects of it will appear in different modules anyway. Students will get exposed to small case studies throughout the year. Therefore, a short submodule on uncertainty should be kept for the end of the year as a unifier/highlighter of these different aspects. This discussion aiming at unification can be supported by a case study like Alice. This may as well be a good pointer for bunch of good open research directions.
(2) Bayesian methods and graphical models seem very useful and powerful and they need to be taught. As far as we see the submodule on these topics can be taught any time after covering basics of probability and graph theory.
We talked to Yaser Abu-Mostafa on Bayesian methods and graphical models before the meeting. He says there is a course on campus on this topic and also he teaches the basics in his learning class over about 4-5 lectures. Andreas Krause mentioned his "probabilistic graphical models" course offered next year. He also claimed that if needed these topics can be condensed to about 6 lectures. He also mentioned something called "robust Bayesian theory" and claimed it might the framework to unify different uncertainty models and graphical models. Another claim is that stochastic optimal control can be covered in the intersection of Bayesian theory and graphical models (as well as Kalman filter etc.).
John's claims: Uncertainty as we have in controls (parametric&worst-case) can be merged into stochastic processes course and it would not be a huge jump in the context. This is not necessarily true for unmodeled "dynamics" since such a notion is widely utilized.
General consensus: Graphical models can serve a basis for a lot integration of topics of interest and uncertainty can be handled in different modules with possible unification at the end.
Toward the end of the meeting, people started talking about a rescheduling in the meetings to have room for a presentation by Krause.