Weighted approximate Bayesian computation via Sanov’s theorem

Abstract We consider the problem of sample degeneracy in Approximate Bayesian Computation. It arises when proposed values parameters, once given as input to generative model, rarely lead simulations resembling observed data and are hence discarded. Such “poor” parameter proposals do not contribute at all representation parameter’s posterior distribution. This leads a very large number required and/or waste computational resources, well distortions computed To mitigate this problem, we propose an algorithm, referred Large Deviations Weighted Computation where, via Sanov’s Theorem, strictly positive weights for thus avoiding rejection step altogether. In order derive computable asymptotic approximation from result, adopt information theoretic “method types” formulation method Deviations, restricting our attention models i.i.d. discrete random variables. Finally, experimentally evaluate through proof-of-concept implementation.