Failure Modes of Variational Inference for Decision Making

In this paper we highlight the risks of relying on mean-field variational inference to learn models that are used as simulators for decision making. We study the role of accurate inference for latent variable models in terms of cumulative reward performance. We show how naive mean-field variational inference at test time can lead to poor decisions in basic but fundamental quadratic control problems with continuous actions, as relevant correlations in the latent space are ignored. We then extend these examples to a more complex non-linear scenario with asymmetric costs, where regret is even more significant.

• Machine Intelligence

• Algorithms and Theory