HamiltonianABC.PDFVIP

Hamiltonian ABC Edward Meeds Robert Leenders Max Welling Informatics Institute Informatics Institute Informatics Institute University of Amsterdam University of Amsterdam University of Amsterdam tmeeds@ leenders.robert@ welling.max@ Abstract using as inputs and producing outputs . If the pseudo- data is “close” to observations , then is kept as a sam- Approximate Bayesian computation (ABC) is a ple from the approximate posterior. Parameters are then powerful and elegant framework for performing adjusted, depending upon the algorithm, to obtain the next inference in simulation-based models. However, sample. due to the difficulty in scaling likelihood esti- In ABC, there is a fundamental trade-off between the com- mates, ABC remains useful for relatively low- putation required to obtain independent samples and the dimensional problems. We introduce Hamil- approximation to the true posterior. If the parameter mea- tonian ABC (HABC), a set of likelihood-free suring closeness is too small, then samplers “mix” poorly; algorithms that apply recent advances in scal- on the other hand, if it is too large, then the approxima- ing Bayesian learning using Hamiltonian Monte tion is poor. As the dimension of the parameters grows, the Carlo (HMC) and stochastic gradients. We find problem worsens, just as