cat /proc/claus

The 5th day of GECCO (friday, 10th) was the first day of paper presentations. With only two days of paper presentations (as compared with the three days of GECCO 2008), it felt hard to attend to all the presentations I wanted to. I’ll probably send an e-mail to the GECCO organizing comittee suggesting that they spread the presentations a bit more next year.

Also, since I had to present two works in different sessions, my choices were further limited. Oh well. Anyway, the first session
I attended was “Financial Applications 1″, with two works by Chris Clack, Wei Yan and Ghada Hassan from UCL, which were extensions of their works in the previous edition of GECCO. I particularly like Ghada’s work, which focuses on not allowing a pareto front of solutions to “switch places” (move to a different area of the pareto front) when the environment changes. There was also a work on using developmental solutions for stock trading, which I still have to read more closely (TODO: and refer it to my trading colleagues in the lab).

The second session was the RWA best paper, where I had to present the work on generating trading rules with GP by my labmate Hirabayashi. Unfortunately, as Murphy would have it, my laptop and the projector decided that they don’t really like each other, and would cut communication between each other every few minutes (even though I checked everything and it was alright just one hour before :-(). Oh well. After the presentation, I made the acquittance of Stefan Haflidason, a very pleasant fellow from the UK. We had lunch together and talked about his plans to do a Post Doc in Japan.

In the afternoon, I wanted at first to go to the “GECCO Job Shop”, where poor grad students looking for work would supposedly meet with people with open positions… but no one was in the room. Instead, I had the pleasure to attend to the HUMIES award presentation. The HUMIES is a competition of research works on evolutionary computation which are not only effective, but competitive with human habilities in fields which are traditionally dominated by humans. There is a cash prize to the best work, and all the presentations were of very high quality. Among the candidates, the ones I found most interesting was a system that could find and correct bugs in programs - which was demonstrated on outstanding bugs in varied open source projects! Obviously that one won the prize this year. Other works that I liked were a system that generated malware programs, and a system that used IEC to generate drum accompainment to musical pieces. On a more light hearted note, one of the presenters sounded just like Dr. Strangelove - so much that I was half-expecting his hand to grab his throat at any time during the presentation.

The last session I attended to was second financial applications session, where I presented my own work on Memetic Algorithms for Portfolio Optimization. This time, the laptop and the projector respected each other, and the presentation ran without a glitch. I had a number of questions, and it was a very fine experience.

After the day was done, I tried to get together a largish group of people to go and walk around Mt. Royal - but ran into a complicated constraint satisfaction problem. A group of Romanians wanted to go walking, a group of Japanese wanted to go by train, and everyone had different ideas about how hungry we were. In the end, we separated our 15 people group in two: one to go walking, and one to go by train, and decided to meet by the entrance of the park. By a freak accident of destiny the walking group got to the meeting point before the group that would go by train, which led me to believe that we missed the meeting point. Me and Stefan tried to find the train group, and in doing so, we missed them completely (they got to the meeting point 5 minutes after us) - when we returned, the Romanians had returned home, some people already had dinner, and nobody wanted to go to Mt. Royal anymore :-P. Me, Stefan and about 10 Japanese went to a nearby American-style diner, and managed to salvage the night with some nice lasagnas. After so much walking, I returned home and slept like a stone.

The next day, Saturday, was a bit overcast. The first session I attended was on “Dynamic Environments and Aging” - and it had two papers which were quite relevant to my work. One was the use of ALPS (age-layered population structure) to maintain diversity in the population for Real-Valued problems (by Gregory Hornby). The basic idea of ALPS is to give each individual in a GA an “age” (based on the evolutionary operators it went through), and separate the population into age layers, so that only individuals in the same layer could mate to each other. One interesting point in this particular work is that older individuals tended to be more complex and solve the problem better, and it was suggested during the discussion stage that individuals in different age layers could be trained in different, and progressively difficult, problems, to achieve some sort of developmental property. I found this idea really cool. The second paper which I found relevant was about “Terrain-Based Memetic Algorithms” (TBMA) by Carlos Azevedo and Scott Gordon (couldn’t find any links handy, sorry). The basic idea was to generate a grid, and distribute the individuals of the population in the grid - individuals could only mate with other geographically near individuals. Then for each position in the grid, you assign different parameters for the local search of the evolutionary algorithm. In this way, the movements of the individuals over the grid as they mate and are selected by the fitness function will lead to a self-adaptation of the best LS parameters (the grid locations with the best parameters will attract individuals). I found that this idea can be expanded in a very neat way to my project (and I still have to contact the authors to talk about this.)

After the DEA session, I attended to the game session where I saw Omid’s wonderful presentation about his computer chess player. He has this very simple and elegant idea to train chess positions evaluators using a database of movements from a human grandmaster. In a way, he is reverse engineering the mind of Kasparov. The impressive thing about his current work is that training multiple times on the same dataset, he got a number of chess players with different characteristics. Since he reported that none of his chess players managed to get more than 40% of the answers right, I’m guessing that the difference comes from learning different parts of the database that correspond to different styles of play from the same player. So I guess it is not even possible to go much above this threshold, since the “evaluation” function used by the grandmaster may have been actually quite different for different groups of moves. It is a fascinating work and I can only guess at what Omid will come up with next year.

For lunch, I went back to lovely Rue St. Denis, and to the “Valet du Cour” game shop, to buy a deck of cards that Leon had asked me. I also bough myself a plush green dragon - I couldn’t resist the little beast. It started raining, and I took shelter at local Cafe. All the waiters could speak mostly French with some pidgin english, and I felt a bit ashamed for not being able to communicate properly with them. Need to learn more languages - or at least make an effort to learn the basics out of a phrasebook before visiting other countries.

Because of the rain, I didn’t make it back in time for the third session in that afternoon. I arrived a few minutes before the end of the session, and while I was browsing the net in the lobby, it was curious to observe the network getting considerably slower when the session ended and the influx of geeks turning on their laptops began.

After that we had the plenary session, with a talk by professor Demetri Terzopolous - It did not have anything to do with evolutionary algorithms, but I guess that is kinda the point of a plenary session, right? Demetri talked about is works on the simulation of individuals and societies, describing the framework for realistic simulation of complex creatures, from the physics level to the structural, muscular, reactive and behavioral levels. Plenty of pretty pictures and animations too (Although it got a bit awkward when he showed that strip scene from Animatrix).

At night, we had the main reception at the Montreal Science Center. It was supposed to be GECCO’s 10th birthday, so we had a cake and fireworks (not provided by SIGEVO *g*). It was a good chance to shake hands and talk to many people I had not yet had the chance to meet. In particular, I was introduced to Nuno Horta and his student Antonio Gorgulho, who used my master thesis in some of their earlier works - It may be kinda silly, but it felt nice to know of someone who studied your works!

After the party, I went with Mark and Jeff and a few others for some extra drinks in a bar in old town. It rained quite heavily while we were looking, and none of us had umbrellas, so when we finally picked a place, we were quite drenched. I didn’t stay for long, for I was quite tired, but I heard they had a terrific time there.

And that’s for 5 and 6. Sorry for the delay! I’ll try to get the last part of my travel log online soon, so we can get on with our regular schedule! (I have postponed some stuff I wanted to talk about until I could get these GECCO posts finished).

GECCO itself began on my third day (second full day) of my trip to Montreal. The first two days of the conference were to be tutorials and workshops, and the last three days were the paper presentations.

The registration went smoothly, although I found the bonus toy lacking: it was a weird gadget with an extendable ruler, an horizontal bubble, and a pad of sticker notes. Huh? A bag would have been much more useful. The venue chosen for the conference, on the other hand, was very good: A nice hotel, with very efficient servers and nice conferencing rooms.

The first session was a financial tutorial by Christopher Clack, from UCL (UK). Christopher made a very nice overview of the internal workings of a financial company, highlighting the many different areas which we of Artificial Learning background can contribute with. After the talk, I had the pleasure of meeting Ghada Hassan and Omid David, two people I have met in GECCO 2008, in Atlanta. I was very happy that both of them recognized me as well right away, and we spent a whole lot of time talking during the conference.

The second talk I heard was from Natalio Krasnogor, about the simulation of Biological systems. At first I was interested to see if he would link any of his work on Memetic Algorithms, which I had read a lot these past few months. He didn’t, but it was an interesting talk nonetheless.

During the afternoon, I attended the “Failures in Evolutionary Computation” Workshop during the afternoon. The title and the summary of the workshop attracted me. I have this notion that ample divulgation of failures is a very important in any scientific field, and that the reality of “publish and perish” causes a bad situation where any study which is not a “complete success” in some form is thrown away and forgotten. And in this way, we probably have many groups re-inventing the wheel all the time. So I was expecting the workshop to be a step in the opposite direction, where participants would describe some ideas that they had that were spectacularly wrong, and how they eventually found out about the error in their ways.

However, it seems that old habits die hard, and most of the papers in the workshop were more of the sort “I tried this method to solve the problem, and as you can see it was not very successful. But then I tried to change the parameters a little, and it worked really well!”. One of the invited talks was interesting, the speaker talked about how many bad individuals in a population would be necessary to guarantee good results in ES techniques. But even then I was a bit disappointed with the tutorial. I certainly hope that next year people will be more bold to talk about their mistakes - I’ll try to send a paper about my misfortunes with ant clustering.

At night, I went dining in a very nice French restaurant. I fount it a bit funny that, for about the same price, the beef/seafood dishes were so tiny (as one would imagine an stereotypical expensive French restaurant), while the pasta dishes were huge. There was the GECCO student party at a student house in the local university, but while I attended, I was so tired from Jet Lag that I soon returned home.

In the second day, there weren’t any tutorials for which I had a strong desire to attend to. In the morning I saw a tutorial about Generational and Developmental systems. Developmental systems were a big subject during GECCO 2008, and basically means evolutionary systems where there is a complex process which leads from the genotype to the phenotype, and this process can be influenced by the environment. The cool thing about this tutorial was that it had a very large bibliography about GDS, it gave a very broad overview that I want to make sure to read on once I have some time for such.

The next talk was about techniques for running experiments with Genetic Programming. The “meat” of the talk was the use
of programs which automatized the analysis of the parameters in an experiment. Then in the next time slot, I got into the wrong room by mistake, and entered the end of a Workshop on Symbolic Regression. It was a surprisingly good mistake, because the papers in that workshop were all very oriented to practical applications, and interesting.

The end of the workshop day was the poster Session. The poster session is always one of my favorite parts of a conference. You get a chance to talk to many different people in a short span of time, and get many different ideas thrown at you. It has a brainstormy feel to it that I really like. For this year’s edition, a crossover operator for index arrays and a policy to dynamically control the number of offspring generated by crossover every generation were the things that stood on my mind the most.

After the Poster session, I got together with a group of latinos and a group of romenians, and we went all together to the Jazz festival. I mentioned to the latinos that I heard about a guy playing salsa at the festival that night, and they beelined to there like a shark smelling blood. Learned that the Romenian language has some similarities to Portuguese.

I’m getting absolutely tired of this computer screen. For the past month I have been trying to write an article for a specific journal at the request of my professor. The biggest problem is that I JUST had a paper accepted in another journal, so I didn’t really have any new results to publish. I had to set up new experiments in a hurry to test for some theories and methods we were thinking about, but everything was rushed and slapped together with spit and glue.

At least, the deadline is two days ahead of me - I only have about 60% of the experiments completed. I have spend most of the previous week writing the “meat” of the paper and trying to fix the experiments, but by now I have mostly given up, and I’m just trying to write a paper which is good enough not to be thrown out at the first round of reviews, and do the real fixing in the second submission. This is extremely demoralizing.

At the moment I have finished all the coding that I’m willing to do, and run the experiments and see what comes out. I can’t wait to take my life back into my hands; so much to do which is being wasted away by the stress with this paper :-/

Today Professor Lee Spector, , a well known researcher in the field of Evolutionary Computation, came to visit Iba Laboratory. The plan was to give him a demonstration of our research work, then listen to a talk from him, and then a dinner at a nearby traditional restaurant.

The whole day was delightful, and very inspiring. Professor Lee has what the japanese call “playful heart” (遊び心), which means that he really seems to take joy in letting his imagination roam free in his ideas. Talking to him reminded me of how fun it is to do research about GA/GP - taking a computer and making it learn things in a totally emergent fashion. One part that rekindled my interest was when he talked about “getting people from other fields to share a beer with you”, to get them to tell you their problems, so you could try to solve them with GP. It reminds me of what I used to think about CS, in which it is a tool for other sciences to achieve nice things, and not an end to itself - another lesson which I need to remind myself of. It was also quite a Feynman-ish thing to say. The technical contents of his talk were two very simple GP ideas that could be easily implemented and tested in our systems - his main point was that GP needs more simple ideas than extra complex systems.

I got to play tour guide and take his family around the University of Tokyo, and that was really refreshing. I was surprised about how much they knew about his work - It got myself thinking that this was the kind of familiar/professional position I would like to see myself 10-15 years from now.

Now to go back to my papers and start working towards this direction

While my official graduation estimate is October 2010 (15 months from now), I have been recently informed that I actually have to finish the most important parts of my thesis by January (7 months from now). This news fell on me like a sack of bricks, with the newfound realization that I have to seriously step up my research efforts (I can only hope this also means that I get a more relaxed pace after January).

This is as great a time as ever to get stumped in my research. I just reached a snag that I have been trying to think my way out of the whole week, without much success. Here is the story:

So far, I have developed a novel computational method to the “Portfolio Optimization Problem”, which is financial problem, and also a specific subclass of the Parameter Optimization Problem. Now, my method has achieved results which are much better than anything currently out there, and I managed to produce quite a few good publications from these experiments. But that can only go so far.

The next step would either be to focus myself more on the Portfolio Optimization Problem, or to try and generalize my method to a broader category of Parameter Optimization Problems. The first, while a worthwhile area of research, is not exactly what I find myself doing for the rest of my life - the financial market does not attract that much of my interest. If I were to follow this, I would have to study economy to a degree I don’t really think I’m willing to put up with.

The second option is more the kind of thing that I want: Theoretical research. If I can modify my algorithm to solve a wider variety of problems efficiently, I can maybe learn something about the nature of this category of problems, or of algorithms like mine, and talk about that. This kind of general, theoretic research which makes my brain jog and exercise is what I really like to think of myself as doing. So I’m trying to follow the second path.

Problem is, it seems that my method has its good results because it is REALLY fine tuned to the specific characteristics of the Portfolio Optimization Problem. I’m trying as hard as I can to generalize the algorithms to problems that don’t share some/all of those characteristics/restrictions, and although I had a few good ideas, all methods I have come up with so far have been too complicated and inelegant. Normally, I could just try to scrape this method completely and try anew, but the fact that I have only 7 months to prepare and write my thesis, and that I already have 21 months of work in my previous method that I should be using on it makes this alternative not very attractive.

So either I come up with some great idea in the next two weeks or so, or I’ll gave to go down the first path, at least for my phd, and hope I can turn around in time to do some work I like in my post-doc/profession.

At times like this I would like to still have my old professors, LM and Wainer, with whom I could sit a whole hour, bouncing ideas back and forth. Well, I have a lab meeting next Tuesday, where I am suppose to present my research progress… That will have to do. Hopefully I can come up with enough silly ideas until them so that I’ll be able to spark a brainstorm session during the meeting.

(I know I have been deliberately vague with the details of my work - this is what you have when you’re in the academic world with a paper down the pipeline. Mail me if you are interested in details.)