[Pearl;
Lauritzen & Spiegelhalter; Jensen et al., Dechter; Amir; …]
Our
language exploits locality:
nAll information is
presented in terms of local influences
Locality
also allows divide & conquer inference
nIn Bayesian
networks
nIn constraint
satisfaction
nIn logical reasoning
CS101
Stat116
Geo101
Stat116
Stat116
Divide and conquer inference
allows us to perform inference separately over different parts of the
knowledge base, and to combine results by sending a “message” involving only
a small number of entities, e.g. (in this simple case), those in the
intersection between the parts.In
general, this extends to hierarchical structures.