This is the code package for the article

Antti Hyttinen, Frederick Eberhardt, Matti Järvisalo:
"Constraint-Based Causal Discovery: Conflict Resolution with Answer Set Programming"

published in Uncertainty in Artificial Intelligence 2014, Quebec City, QC, Canada.

1. Fetching Clingo:
----------------
This code package does not include the clingo ASP-solver. Clingo is available
for mac,linux and windows:

Fetch it from address

http://sourceforge.net/projects/potassco/files/clingo/4.3.0/

and replace the empty file "ASP/clingo430" with the file.

We use clingo 4.3.0 mac version. The code has only been tested on mac. The R code
calls clingo automatically.

2. Loading the code:
-----------------
Start R in the "R/" directory and load the code with the following commands:
> source('load.R')
> loud()

Make sure you have packages 'deal','pcalg','combinat','graph','hash' and 'bnlearn' installed. The code was produced using Rstudio and R version 3.0.2 (2013-09-25).

3. Running simple tests:
---------------------

>test(1)
>test(2)
>test(3)
>test(4)

4. Producing similar figures as in the paper:
-----------------------------------
First run the algorithms to produce the data in the directories.
> plot1(run=TRUE)
Then produce the plot.
> plot1()
Similarly for plot1....,plot7. See the R-files for more information.

5. What is still missing or not implemented well enough:
-----------------------------------------------------

-If you fail or interrupt a run some of the files that the program writes constraints into may remain open for writing and some output meant for the screen will be writer to those files. Then next run fails. Typing “sink()” a couple of times will remove this.

-In principle the code should handle experimental overlapping data sets, all of
the theory surely allows for it. However this code package has not been excessively
tested in this scenario. Also the scalability may suffer at least from the use of several data sets.

-The code basically outputs only one graph which maximizes/minimizes the objective
function. To get an idea what features are actually determined one should run HHEJ 2013 algorithm with the single graph of this paper as an input. There might be a possibility to build the same functionality over clingo, but it is not sure whether it would match the efficiency of the HHEJ 2013 paper method. For
restricted setting one might run for example FCI on the d-sep oracle with our output to produce a PAG.
Yet another option would be to print out all solutions from Clingo (see instructions of Clingo) and
see the common properties there.

6. Disclaimer on the ASP encodings:
-----------------------------------

For ease of reading the UAI paper presented the encoding with sets
 V (variables in the d-connection graph), C (conditioned vars) and J (intervened
 vars).  The code which was done earlier considers
  C ( conditioned vars), J (intervened variables) and M (marginalized variables). 
The notations are equivalent, V of the paper notation is the set of original variables
minus C and M. 

In addition the d-connection graphs permitted self-cycles between variables. As only undirected self loops x—-x may affect d-connection properties of the graph, only these are considered in the code. For example a path y—>x<—>x<-z implies the existence of a path y->x<-z, which is d-connecting for exactly the same conditioning sets, so x<->x can always be ignored.

6. Questions
------------
Please try to figure out yourself especially if the problem is in installing or getting the code running. Otherwise email:

Antti Hyttinen <[email protected]>