Tractable Bounding of Counterfactual Queries by Knowledge Compilation
Explore a 34-minute lecture on bounding partially identifiable queries in Pearlian structural causal models. Delve into an iterated EM scheme that yields inner approximations of bounds through parameter sampling. Examine how compiling the underlying model to an arithmetic circuit can significantly speed up inference for models with shared structural equations and topology but different exogenous probabilities. Learn about symbolic knowledge compilation techniques that allow for flexible parameter substitution in circuit structures. Discover parallelization methods to further accelerate bound computation. Compare the computational advantages of this approach against standard Bayesian network inference, revealing up to tenfold speed improvements.