This repository will contains for "Approximate Knowledge Compilation by Online Collapsed Importance Sampling", as presented at NeurIPS 2018.
Requirements: Scala SBT Jython standalone jar
Most of the necessary preprocessing is handled by run_pipeline. It requires a single argument (the uai file), and will output a directory that looks like DBN_11_processed. The source of run_pipeline.sh should be modified with the appropriate location for your jython jar.
50-20-5_processed is an example of an experiment that has a logical base which we use. The files 50-20-5.sdd and 50-20-5.vtree.out were created by the SDD tool available at http://reasoning.cs.ucla.edu/sdd/.
Some example experiments from the paper:
For experiments that don't use a logical base:
export LD_LIBRARY_PATH=./lib/ && sbt "run -l DBN_11_processed/DBN_11.litmap -u DBN_11_processed/DBN_11.uai -v DBN_11_processed/DBN_11.vtree -d DBN_11_processed/DBN_11sdd/ -q -1 -t 3600 -s 1000000 -H rbvar -O rev -b DBN_11_processed/DBN_11.base.cnf -e DBN_11_processed/DBN_11.uai.evid"
This uses the rao-blackwell variance minimizing heuristic, reverse ordering, a size limit of 100,000 nodes, the last variable as the query, and runs for 1 hour.
For experiments that do use a logical base: export LD_LIBRARY_PATH=./lib/ && sbt "run -l 50-20-5_processed/50-20-5.litmap -u 50-20-5_processed/50-20-5.uai -v 50-20-5_processed/50-20-5.vtree -d 50-20-5_processed/50-20-5sdd/ -q -1 -t 600 -s 1000000 -H fd -O rev -B 50-20-5_processed/50-20-5.sdd -V 50-20-5_processed/50-20-5.vtree.out -M -E -e 50-20-5_processed/50-20-5'.uai.evid'
This uses the frontier distance heuristic, reverse ordering, a size limit of 100,000 nodes, the last variable as the query, and runs for 10 minutes. It also uses a precompiled oracle for the logical base, and applies entailment checking after conditioning.
If you have any questions, please drop me an email at [email protected].