Improve approximate inference algorithm

For problems intractable with exact inference, Lea provides a method that performs approximate inference, namely estimate_mc. This works fine but, because it is a rejection sampling algorithm, it has the drawback to be inefficient for calculating conditional probabilities x.given(e) when P(e) is small: most of the calculation time is passed to generate, test and drop random samples not verifying the condition e.

Lea should provide other approximate algorithms, which are more efficient for this kind of problems. MCMC and weighted algorithms are not directly envisioned here, because AFAIK these are limited to "observations" (conjunction of equalities) while Lea is more general, covering any boolean function for the condition e. An "hybrid" algorithm mixing exact inference on e and approximate inference on x should be feasible. Also, extra options beyond the number of samples shall allow to make trade-offs in those algorithms.

Comments (8)