≈<? require("header.html"); ?>

<h1>THE SOLVERS OF THE 2010 UAI APPROXIMATE INFERENCE CHALLENGE</h1>

<ul>

<li>
<font size=5>IJGP: UCI<br/></font>
<b>Summary</b>: Iterative Join Graph Propagation (IJGP) is a generalized belief propagation algorithm that performs the same message passing as join tree propagation, on a polynomially bounded join graph, iteratively.<br/>
<b>Team members </b>: Vibhav Gogate, and Rina Dechter<br/>
<b>Additional contributors</b>: Kalev Kask and Robert Mateescu <br/>
<b>Additional Support</b>: Andrew Gelfand <br/>
<b>Contact person</b>: Vibhav Gogate, vgogate@cs.washington.edu<br/>
<a href="docs/vibhav-ijgp.pdf">UAI Presentation Slides</a>

<p><li>
<font size=5>FormulaSampleSearch: UW and UCI<br/></font>
<b>Summary</b>: In formula-based importance sampling, we perform importance sampling over the search space defined by truth-assignments to logical formulas instead of the variables. Since, the formula-based search space is provably smaller and more informative, formula-based importance sampling has smaller variance than variable-based importance sampling.<br/>
<b>Team members</b>: Vibhav Gogate, Pedro Domingos, and Rina Dechter<br/>
<b>Additional Support</b>: Andrew Gelfand <br/>
<b>Contact person</b>: Vibhav Gogate, vgogate@cs.washington.edu<br/>
<a href="docs/vibhav-formula-samplesearch.pdf">UAI Presentation Slides</a>

<p><li>
<font size=5>ED-BP: UCLA<br/></font>
<b>Summary</b>: An anytime solver that starts with a
tree-structured approximation corresponding to loopy BP, and
incrementally improves it by recovering deleted edges.<br/>
<b>Members</b>: Arthur Choi and Adnan Darwiche, with support from Glen
Lenker and Knot Pipatsrisawat.<br/>
<b>Contact person</b>: Arthur Choi, aychoi@cs.ucla.edu<br/>
<a href="docs/ArthurChoi.pdf">UAI Presentation Slides</a>

<p><li>
<font size=5>libDai: Max Planck Institute<br/></font>
<b>Summary</b>: Combination of algorithms: junction tree, loopy belief
propagation, a double-loop GBP algorithm, a Gibbs sampler, the
max-product algorithm and a decimation algorithm.<br/>
<b>Team members</b>: Joris Mooij<br/>
<b>Contact person</b>: Joris Mooij, joris.mooij@tuebingen.mpg.de<br/>
<a href="docs/JorisMooij.pdf">UAI Presentation Slides</a><br/>
<b>Code</b>: available as part of <a href="http://www.libdai.org">libDAI</a> 0.2.6

<p><li>
<font size=5>Toulbar2: INRA<br/></font>
<b>Summary</b>: Toulbar2 is an open source exact anytime Weighted CSP solver using
Branch and Bound and soft local consistency<br/>
<b>Team members</b>: S. de Givry, D. Allouche, A. Favier, T. Schiex<br/>
<b>Additional contributors</b>: M. Sanchez, S. Bouveret, H. Fargier, F. Heras, P. Jegou, J. Larrosa, K. L. Leung, S. N'diaye, E. Rollon, C. Terrioux, G. Verfaillie, M. Zytnicki<br/>
<b>Contact person</b>: Thomas Schiex, Thomas.Schiex@toulouse.inra.fr<br/>
<a href="docs/ThomasSchiex.pdf">Detailed description</a>

<p><li>
<font size=5>Daoopt: UCI Irvine<br/></font>
<b>Summary</b>: "daoopt" and "daoopt.anytime" are based on AND/OR branch and bound graph search, with mini bucket heuristics and LDS (Limited Discrepancy Search) initialization.<br/>
<b>Team members</b>: Lars Otten, Rina Dechter <br/>
<b>Additional Contributor:</b> Radu Marinescu<br/>
<b>Contact person</b>: Lars Otten, lotten@ics.uci.edu<br/>
<a href="docs/daoopt10.pdf">Detailed description</a><br/>
<b>Web-site</b>: <a href="http://graphmod.ics.uci.edu">http://graphmod.ics.uci.edu</a>

<p><li>
<font size=5>Alphabet SOUP: Stanford<br/></font>
<b>Summary:</b> The inner-loop of the algorithm uses
max-product belief propagation, and the alphabet restrictions
are chosen randomly, i.e., the code does not exploit any special
structure of the problem.<br/>
<b>Team members</b>: Stephen Gould, sgould@cs.stanford.edu<br/>
<b>Paper</b>: <a href="http://www.stanford.edu/~sgould/papers/cvpr09-alphabetSOUP.pdf">http://www.stanford.edu/~sgould/papers/cvpr09-alphabetSOUP.pdf</a><br/>
<b>Code</b>: code available as part of the <a href="http://ai.stanford.edu/~sgould/svl">STAIR Vision Library</a>



</ul>

<? require("footer.html"); ?>