Today in HEC 438, 2p-3p, Adam

September 26, 2008
2:00 pm	to	3:00 pm

Today (Fri, 26.Sep), Adam will talk about his recent work at our joint NCCS lab and EC lab meeting.  As usual, we meet in HEC 438 from 2p - 3p.

I will also bring the PPSN proceedings with me, as well as my annotated program.

This semester’s schedule

October 3, 2008
2:00 pm	to	3:00 pm
October 10, 2008
2:00 pm	to	3:00 pm
October 17, 2008
2:00 pm	to	3:00 pm
October 24, 2008
2:00 pm	to	3:00 pm
October 31, 2008
2:00 pm	to	3:00 pm
November 7, 2008
2:00 pm	to	3:00 pm
November 21, 2008
2:00 pm	to	3:00 pm
December 5, 2008
2:00 pm	to	3:00 pm

Here are the meetings we have currently scheduled for this semester.

• 29.Aug Organizational meeting
• 5.Sep Chris presents
• 12.Sep AI Forum
• 19.Sep Round table (canceled)
• 26.Sep Adam presents
• 3.Oct AI Forum
• 10.Oct Sean presents
• 17.Oct Stephen presents
• 24.Oct Ozlem presents
• 31.Oct TBA
• 7.Nov AI Forum
• 14.Nov Gautham presents
• 21.Nov TBA
• 5.Dec TBA

PPSN X Debrief

As most of you know, I recently attended to 10th International Conference on Parallel Problem Solving from Nature (PPSN X).  PPSN is a mid-sized, “cis-Atlantic”, good quality conference for all manner of nature-inspired computational mechanisms.

This year,  papers involving CMA or addressing multi-objective problems in some way represented the majority of work.  I have the proceedings in my office, though they are also available on-line via Springer.

I tried to jot down articles that I thought might be interesting to various people in ECL and NCCS lab.  I have such annotations for Gautham, Yinghua, Dr. Stanley, YanYan, and Adam.  If you care, stop by my office or email me and I’ll try to list the papers in which I thought you may have an interest.

Here were some of my personal highlights:

• There was a very nice poster by André Neubaur that took the Vose dynamical systems model and considered a slight variant of the standard GA to develop a model of this algorithm that could be separated by a pair of commutative permutations in such a way that the basic map was entirely independent of the fitness landscape and selection.  Since there is only one intrinsic system, he could solve for the stable fixed point and determine a number of interesting properties about the system regardless of the problem landscape.  He even used his ideas to connect some of Vose’s ideas on coarse-graining to Schema theory — and applied a real algorithm to validate some ideas.  I’m eager to read the paper.
• Olivier Teytaud developed the ideas he discussed at Dagstuhl on using VC-dimension to establish lower bounds for rank-based search methods.  I’ve read the paper, and though a bit technical, the idea is quite clever (I think).  It’s worth the read.
• Frank Neubaur won best paper for his work on analysis several EMOO algorithms that approximate minimal multicuts in graphs by reversing the idea of “flow”.  I have not read Frank’s paper yet.  I was more interested in his paper with Dimo and Tobias on hypervolume indicator based algorithms.  It gives me ideas…
• Christina Zarges did a great paper analyzing simple variants of artificial immune system algorithms on simple problems.  She examined a particular parameter that the AIS community (apparently) have found very difficult to “get right” and explains pretty clearly why they may have difficult getting their heads around it.  This paper was one of the two that shared the best student paper.  I read it the night before I left … great paper.
• The other student paper award went to a paper by one of Darrell Whitley’s students (Monte Lunacek) for a nice empirical look at some types of structural aspects of landscapes that can trip up CMAs.  His problem landscape looks a lot like MTQ, and their experimental factors were quite similar. Perhaps I should take a closer look?
• Jeff Clune had a  paper looking at the impact of irregularity on the performance of HyperNEAT.  I found his notion of “irregularity” a bit extreme, but I was very happy see some constructive empirical work examining HyperNEAT’s power and limitations.  His method seemed careful enough (from the poster presentation).
• Neal Richter had a paper where he took the Storch & Wegener problem for which crossover is provably essential and modified it to show a problem where crossover was provably harmful.  I had the opportunity to present that poster for Neal, which was actually quite fun.  I was very happy to see another dynamical systems person approach runtime analysis.
• There were also a number of niching / fitness sharing papers, including one that applied some of the methods similar to mutation adaptation to adapt the niching parameter.
• Jeff Horn finished up his triad on analyzing when fitness-sharing methods converge to “complete coverage”.  His paper is straightforward algebra … very easy to follow and very useful (I think).
• Also, there was one by Evelyne Lutton’s group where they used fitness sharing (among other things) to try to predict the phases of cheese ripening as well as experts based on measured properties of the cheese. One can only hope that no one will ever be deprived of making their living tasting cheese …that would be unconscionable.
• There was a self-play RL poster that I should read, though I am a bit skeptical based on the presentation.  The paper examined when RL via self-play fails and attempts, but my gut tells me the properties they identify are secondary, not causal.
• One of Tina Yu’s students and she had a cooperative coevolutionary paper that employs an interesting way to distribute fitness evaluation
• There was another coevolution paper involving CAs and chemical computing, but I missed the poster presentation and have not yet looked at the paper.
• And Edwin de Jong’s group apparently has a coevolution paper in PPSN, though I do not believe anyone was there to present the paper.

CMS Upgraded on NCCS lab website

I upgraded the content management system we use for the NCCS lab website this morning.  I hope there will be no problems.  Please let me know if there are any.

CANCELED: HEC 438, 2p-3p EC lab / NCCS lab round table discussion

Today’s meeting has been canceled.

 EC lab and NCCS lab will hold their weekly joint meeting today in room HEC 438 from 2p-3p.  We will have a “round table discussion”. If people are interested, I can summarize some of the PPSN papers (here is the program). There seemed to be papers that various people in our two groups may find interesting … I annotated my hard copy of the program to mark these when it occurred to me to do so.  The two biggest issues at PPSN this time were CMA and MOO.  There were quite a number of run-time analysis papers, especially on MOO algorithms. 

Today Chris Ellis presents “Actuation Constraints and Artificial Physics Control”

September 5, 2008
2:00 pm	to	3:00 pm

We have a joint NCCSlab / EClab meeting today (Fri. 5.Sep.08) in room HEC 438 from 2p - 3p.  Chris will present a draft of the poster he and I have been putting together for PPSN next week:  ”Actuation Constraints and Artificial Physics Control”.  The PDF of the paper is available here (about 200k), the draft version of the poster is available here (about 1.4mb).  The abstract is below the cut.

Also:  Anyone who plans to come and has access to a laptop, would you mind if we used it for the talk?  Email me or Chris to let us know.  Thanks.

(more…)

Joint EClab / NCCSlab meeting Friday, 2p-3p in HEC 438

August 29, 2008
2:00 pm	to	3:00 pm

The Evolutionary Computation lab and the Natural Computation & Coadaptive Systems lab will hold their first joint group meeting this Friday, 8/29 fro 2p-3p in HEC 438.  The purpose will be to determine the agenda for the semester.

GECCO 2008 Debrief

The 2008 Genetic and Evolutionary Computation Conference was held in Atlanta this week.  The program is available on their website, and the papers should all soon be in the ACM digital library.  If you want access to a paper before then, come and see me.  I will have the CD in my office next week. Highlights:

• The new combinatorial optimization track was very good
• Carsten Witt gave a new tutorial on theory of randomized search heuristics for combinatorial optimization (unfortunately, I missed it, but Carsten does excellent work so I am certain it was good) 
• The theory track papers were also very good (I especially liked Rowe et al.’s work on precision & local search), though theory papers were quite scattered in other tracks
• Most coevolution papers were good (Travis Service’s NFL talk was particularly interesting), though again, the number of coev papers in the actual track was very small
• The competitions track was interesting; there was a lot of discussion
• PZ Myers gave a nice, high-level talk about EvoDevo (I missed the other invited speaker)

 Other observations:

• GECCO 2009 in is Montreal
• Most of the GECCO 2009 contests have already announced (it’s not posted yet, but I may have a flyer)
• There’s no coevolution track next year
• The “Formal Theory” track will be simply called “Theory” next year

The new issue of SIGEvolution is out

You can find the latest issue of the ACM/SIGEVO newsletter here.

2D Packing GECCO Contest Submission

Gautham Anil, Adam Campbell, Chris Ellis, and Yinghua Hu put a lot of work into their joint attempts to find a good solution for the 2D Packing contest at GECCO this year.  Ours was not the top scoring result, but it was very competitive.  I think congratulations is due to the students for putting in such effort above and beyond their normal school and research duties.  I hope they all agree with me that it was definitely a team venture and a rewarding experience. Indeed, I was most impressed by how the group worked:  Everyone worked on their own, but in our weekly meetings ideas were exchanged.  It was clear that everyone was impacting everyone else’s efforts.  In the end, our best scoring approach (implemented and run by Gautham) really did have a little bit of everyone’s ideas in it (seed & greedy growth from Yinghua’s work, Adam’s shaping ideas, and a variant of Chris’s “smart mutation” work).  Moreover, we all spent some time analyzing the problem and discussing our respective views.  If I learn more about the background of the problem, and the approaches by other contestants, I’ll post them on the website or email them around.  I think we all learned a lot.   Good work, team.  I hope we can put something together next year.