Adap’skiii [day 1]

Adap’skii started on Monday morning with a highly pedagogical and smooth entry by Jeff Rosenthal, followed by a state-of-the-art talk by Eric Moulines on adaptive and interacting MCMC algorithms. Both talks led me to ponder how far adaptivity could reasonably be extended, namely how much the parameters used in those algorithms could themselves be tuned and adapted (as e.g. the temperatures in annealing or tempering). This pondering was actually reproduced later in the day by Nando de Freitas and David Dunson. The  third talk of the morning was Radu Craiu’s adaptive construction of a mixture approximation, including a proof of convergence of the algorithm, which was a bit reminiscent of our population Monte Carlo construct.  After a great sunny afternoon skiing in the [Brad’s correction: balmy at first, bitter once the sun set] cold of the Canyons ski pistes, we were back listening to Nando de Freitas‘ interesting mix of optimisation and MCMC adaptation, using Gaussian processes to design objective function evaluations. The final talk was by Faming Liang about likelihoods with unknown constants and the adaptive estimation of the constant. Pierre Jacob led the discussion presented below: