Once we recast the robust inference problem as the estimation of parameters
Theta, we can also assume the parameters Theta to be assigned
uniform priors. In this case,
any value of Theta that maximizes p(x_{q} = a, Theta|e) also maximizes
p(x_{q} = a|e, Theta). We are interested in algorithms that produce
such maximizing values of Theta, since
_{q} = a|e) = _{Theta} p(x_{q} = a|e, Theta).
With this maneuver, we can use Bayesian learning methods to produce
robust inferences.

Sampling algorithms for calculation of posterior maxima have been studied in connection with Bayesian inference in general [39]. The reasoning in the previous paragraph demonstrates that they can be applied directly to Quasi-Bayesian inferences as well. Simulated annealing can guide a Gibbs sampler in generating samples of the posterior distribution; the sample with the highest probability defines the maximum [39].

This sampling approach offers a contrast between
the interior-point methods advanced here and combinatorial optimization
methods that search for the best combination of transparent variable
values [5]. In combinatorial approaches, each iteration of
the sampling procedure demands a complete cycle of standard Bayesian
inference. Instead, by searching in the interior space of distributions,
we can use the simulated annealing and Gibbs sampling *simultaneously*;
the convergence of this process is a particular benefit of the probabilistic
structure of graphical models [39] which is not exploited
by purely combinatorial approaches [5].

© Fabio Cozman[Send Mail?]

Fri May 30 15:55:18 EDT 1997