Congratulations to Dr Dennis Prangle, who has had a paper accepted at Bayesian Analysis, “Bayesian experimental design without posterior calculations: an adversarial approach”, written with Sophie Harbisher and Colin Gillespie.
The paper proposes a method to speed up Bayesian experimental design and allow it to scale to hundreds of design choices. It replaces the standard way to judge experimental informativeness (Kullback-Leibler divergence from posterior to prior) with an alternative (Fisher divergence). To work well, it shows an adversarial training method is required: an experimenter picks a design while a critic picks parameters to be inferred.
Well done to Dennis on this achievement.
