Christian Robert,
CEREMADE, Université Paris Dauphine & CREST, INSEE, Paris
Title: Adaptive Multiple Importance
Sampling (AMIS)
Abstract :
The
AMIS algorithm consists of 3 steps. In the initialization step a
set of iid uniform points are sampled in the d-dimensional unit box.
The logistic transformation with scale parameter S, is used to bring
the points back to $R^d$. The scale parameter is chosen using the ESS
of the importance weights. Importance sampling estimates of the target
mean and variance are constructed. These estimates are used to
generate, from a d-dimensional Student-T(3df), the initial set of
particles of the second adaptive step. At this stage a temporal
dimension is introduced and global adaptation is performed by an
importance sampling version of Haario et al. (2001) adaptive
Metropolis-Hastings algorithm. To achieve variance reduction an
adaptive version of Owen and Zhou (2000), deterministic mixture
importance sampling is defined. As a byproduct of the mixture and
actualization process we performe on the weights, all particles are on
the same ``weighting scale'' and can be easily and efficiently combined
to get final AMIS estimator. In the final step a
Rao-Blackwellised clustering algorithm is performed on all
generated particles. The number of clusters, K, is chosen via
BIC. IS estimators of mean and covariance matrices on each
cluster are derived. A K-mixture of Student-T (3df) distributions
is used to generate a final set of particles. The
mixture proportions are taken to be the number of particles belonging
to each cluster. The AMIS estimator is obtained by recycling the
particles generated in all 3 steps, with the corresponding importance
weights. The strength of AMIS resides in its completely adaptive
and multi-purpose nature: no tuning parameter is needed and the same
algorithm is proved to perform well on very diverse high-dimensional
target distributions (from banana shaped to mixture with very well
separated modes).
(This is joint work in progress with Jean-Michel Marin and Antonietta
Mira)