Everyone is welcome to attend the next N&O seminar with Kemal Rose with the title 'Algebraic optimization of decision rules'.
The talk will take place in L017. Depending on the circumstances, we will try to also broadcast the talk via zoom. For more information and registration to get the Zoom link via e-mail, please contact Willem Feijen (willem.feijen at cwi.nl), Samarth Tiwari (samarth.tiwari at cwi.nl) or Sven Polak (sven.polak at cwi.nl).
Abstract: Partially observable Markov decision processes (POMDP's) form a broad framework, modelling various real-world sequencial decision processes that are based on partial information. We reformulate reward optimization of POMDP's in terms of quadratically constrained polynomial optimisation. Algebro-geometric tools are used to predict the number of critical points. We conjecture objective value exactness of the first order SDP relaxation wich implies a convex reformulation of the optimization problem.
