The PLP-BN tools are an extensible set of tools designed to relate probabilistic logic programming and Bayesian networks. The PLP-BN tools are written in tabled Logtalk. In order to run them, you need a copy of Logtalk, and to make full use of all its features you also need a backend Prolog compiler which supports tabling. They have been tested using Logtalk 3.91.0 and both SWI-Prolog 9.0.4 and XSB 5.1.

Introduction

The API documentation maintained with lgtdoc can be found in the directory docs; an entity diagram can be found in the directory dot_dias. The documentation also provides an overview of all the features and algorithms included in the PLP-BN tools.

In order to enhance modularity and extensibility and allow for easier porting between Prolog engines, they are written in Logtalk rather than in plain Prolog. One can find a description of Logtalk along with tutorials and handbooks at the official Logtalk website.

Probabilistic logic programs are represented as sets of probabilistic rules of the form probrule(Head,Probability,Body), Facts are stored in a (deductive) database, which provides external predicates for the probabilistic logic program.

The predicates that are to be supplied from a database should be specified in a protocol specific to the probabilistic logic program; see files cora.lgt (contains program and protocol) and cora_db.lgt for a complete example.

Usage

The core tools are loaded by calling {loader}. The supplied examples can be loaded using {example_loader}.

Orienting causal relations

To compute the orientable edges of a graph test1 according to the PC algorithm, call oriented(test1)::edge(X,Y) (returns all orientable edges on backtracking). Symmetries are used whenever they are provided by the input graph; therefore, oriented(graph(uwcse1,uwcse1_db2))::edge(X,Y) orients a smaller number of edges than oriented(graph_psymm(uwcse1,uwcse1_db2))::edge(X,Y), which is able to orient the entire graph.

Transformations

To compute the ground graph arising from a relational probabilistic logic program cora and database cora_db, call graph(cora,cora_db)::edge(X,Y) (returns all edges on backtracking).

To compute the probabilistic logic program in our internal probabilistic rules format from a (simple) Problog program file problog_test.plp contained in your current directory, call plp(problog_dcg,'problog_test.plp')::probrule(Head, Prob, Body) (returns all probabilistic rules on backtracking).

Counterfactuals from a Bayesian network

To compute a probabilistic logic program from a Bayesian network, use the object lewis_cf(BN) for a Bayesian network object BN. This probabilistic logic program has the additional property that its associated counterfactual worlds only differ minimally from the "real world" in a specific sense.

Parametric objects can be combined freely; for instance, oriented(graph(cora,cora_db))::edge(X,Y) orients the edges induced by cora and cora_db.

Interfacing with bnlearn

To support working with Bayesian networks fitted using the popular bnlearn libraries in python and R, the bnlearn_utlities directory contains Python and R functions writing a Bayesian network to file as a Logtalk object. This directory also contains a Prolog/Logtalk convenience script compute_cf.pl to be used as a command line utility for writing a ProbLog program (the input format for e.g. the What If? counterfactual solver) to file that corresponds to an input Bayesian network (as a Logtalk object), using lewis_cf (see above) for the actual computation.

The easiest way to use that script from scratch is to install SWI-Prolog (either from www.swi-prolog.org or from a package manager) and then to call

swipl -g convert compute_cf.pl Input.lgt Output

The first time of calling this script, if Logtalk is not yet installed, the SWI-Prolog interpreter should now prompt you for a directory for installing packs and for confirmation to install Logtalk as a SWI-Prolog pack.

After Logtalk has been successfully installed, conversion should proceed as normal.

If Logtalk has already been installed in the ordinary, backend-agnostic manner as detailed on www.logtalk.org, the script should be launched from Logtalk instead of from plain Prolog.

With SWI-Prolog as the back-end Prolog compiler it can be invoked as

swilgt -g convert compute_cf.pl Input.lgt Output

where Input.lgt is a logtalk object implementing bnp and the resulting ProbLog program is written to Output.plp.

If XSB is used as the back-end Prolog compiler, it can be invoked as

xsblgt -e "[compute_cf], convert('Input.lgt', Output)."

(note the period mark, the " and the ' in the command line flag).