A Python package that implements the k-neighborhood local search algorithms for selecting strategies in (limited memory) influence diagrams described in

D.D. Maua and F.G. Cozman (2014), Speeding Up k-Neighborhood Local Search in Limited Memory Influence Diagrams. In Proceedings of the Seventh European Conference on Probabilistic Graphical Models, LNAI 8754, pp. 334-349.

If you use this code in an academic work, please cite the publication above.

As a by-product, this package also implements a variable elimination scheme for posterior probability computations in Bayesian networks.

Copyright by Denis D. Maua (2014)

THIS CODE IS PROVIDED "AS-IS". USE AT YOUR OWN RISK.

This package is released so that others can reproduce the experiments in the paper, and potentially use the algorithms in other work. You can contact-me with simple questions, but please do not expect to get real support from me.

Simply clone the repository (or download a compressed version). It is highly advisable to run the scripts with the Pypy interpreter instead of the standard Python intepreter. Check pypy.org for info on downloading and installing PyPy.

*** If you have Pypy installed, replace python with pypy in the sequel.

Every runnable script returns its usage if run without arguments (unless the function implemented takes no arguments). Try python spu3.py to get usage on running SPU with dominance pruning.

The diagrams used in the experiments in the paper are in the compressed file experiments.zip. To reproduce the results, uncompress that file and run the script run-chain.py with each algorithm and each diagram, e.g.:

python run-chain.py SPUhull experiments/exp-100-10-?.p

will run SPU (1-neighborhood search) with dominance pruning on ten chain-like diagrams with 100 decision variables and 10-ary variables (each file is a pickle dump of a diagram object).

New chain-like diagrams can be generated with python generate-chain.py. Comparison among different algorithms on randomly generated chain-like diagrams can be obtained running python chainID.py

The codes can deal with arbitrarily shaped diagrams, but there is currently no support for loading an arbitrary model from file (nor there is any standard file format for limited memory influence diagrams). There is however support for loading Bayesian networks in the UAI File Format. With some additional work this can be used to load/save limids from/to file. My suggestion is that the description of node types be given in a separate file (that is, in addition to the .uai file describing variables and CPTs, you should store the labeling of nodes, i.e., decision, chance and value, in a separate file).