Learning Bayesian networks with cops and robbers

The code for the paper "Learning Bayesian networks with cops and robbers" (Topi Talvitie, Pekka Parviainen).

Contents:

• bnrepository contains the networks as downloaded from the bnlearn repository

• bnrepository_nets contains the network structures and CPDAGs preprocessed using preprocess_bnrepository.py (the script requires R and the bnlearn R package)

• tamaki2017 contains the Tamaki-2017 treewidth solver

• This directory contains our code. Usage instructions are given below.

The Makefile compiles all our code and the Tamaki-2017 treewidth solver. Use it by running make. A C++ compiler and a Java installation is required. After compiling, you can use the three resulting executables as follows:

• To test that the algorithm works, run bayesian_test with three arguments: minimum and maximum node counts and time limit in seconds per run. For example, to test it random instances with 0..10 nodes and time limit of 1 second per run, run

  ./bayesian_test 0 10 1
  

  The program will run infinitely (unless it finds an error) and print statistics every 10 minutes.

• To measure independence query count distributions of our algorithm and the PC algorithm when using the exact independence oracle, run bnrepository_test with two arguments: name of the preprocessed network file and time limit per algorithm in seconds. For example, for the alarm network with time limit of 10 minutes, run

  ./bnrepository_test bnrepository_nets/alarm.net 600
  

• To measure the SHD of the learned network from the correct one using our algorithm and the PC algorithm learned from real data, run bnrepository_data_test with two arguments: name of the preprocessed network file and time limit per algorithm in seconds. The program takes the data as input. To generate data from the network, you should use the gen_data.py script (requires R and the bnlearn package). For example, to generate 1000 data points from the alarm network and measure the SHDs of the learned networks, run

  ./gen_data.py bnrepository/alarm.bif.gz 1000 | ./bnrepository_data_test bnrepository_nets/alarm.net 600
  

The code has been configured with a maximum of 128 nodes. To increase this, increase WordCount in bitset.hpp.