cat /proc/claus

In a recent interview, the interviewer asked me to briefly explain my PhD work. While showing him what I did, he intersected “that’s just like Genetic Algorithms, right?” - “Yes, I am actually working on Genetic Algorithms” - “But your curriculum said you worked on “Evolutionary Computation”

…

I can’t really blame him. The field of evolutionary computation (and bioinspired heuristics in general), is plagued by too many “cute/cool” names, such as “Ant Colony Optimization”, “Ant Clustering”, “Memetic Algorithms”, “Queen bee optimization”, “Cat Swarm Optimization”, that often say very little about the technique under them, and less about what other techniques are related.

These names are often the result of well intentioned researchers who want to relate their findings with their inspiration, or who think that a new cool name can bring in more attention, more people and more ideas to the field. But I feel that these names more hinder than help, by making it difficult for people not in the field to have a grasp of what is connected with what, and for people in the field to make sure that what they are doing was not already done by someone else with a different cool name.

Tomorrow I’m taking the plane to Ann Arbor, Michigan, USA, to take part in the “Genetic Programming Theory and Practice” workshop. I’m pretty excited about the opportunity - there will be some of the bigger names in GP theory participating, and the structure of the workshop (few participants, lots of time for presentation and discussion) means that I may actually get the chance to pick their brains for a bit, instead of the normal “big conference” environment where when someone moderately famous makes a talk, everyone rushes to try and talk to them and you are lucky to get a question through.

With some luck, I might even manage to bring up the subject of Post-Doc positions *g* — BTW, I finally got my first serious proposals these days, but it is top secret, don’t bother asking :-).

Logistically, the trip is all set up. I’ll leave tomorrow and return on the 24th — not that it will make a big difference for the pacing of this blog. Recently I have been mostly posting personal stuff to Facebook, and news/research stuff to twitter. I’ll be staying at two couchsurfer’s place — hope the karma pays off :-), and I have printed a small list of Geocaches I want to find during the trip, time permitting. The only part which is a bit uncertain is how I will get from the airport to Ann Arbour — it seems that A2 does not have an international airport (or even a big local one), so I have to get off in Detroit and, this being the US, I can’t take a bus/train from the airport to the city - I need to book a shuttle in advance (which I couldn’t do, since most shuttles only take reservation by phone, none answered my e-mails), or pay 50 dollars for a cab - five times what I would pay for an express train Narita-Tokyo, which is about double the distance. So who said Japan is expensive again?

And don’t get me started on the medieval flight security in US airports…

As an aside, I have finally finished all the experiments and most of the analysis required for the “final” revision of my thesis. I kinda hesitate to use final (hence the “”), since there is a lot of stuff which I’m not satisfied with and would like more time to work on, but it should be enough data to satisfy the defense committee (based on their review of my first defense). So in the next 23 days until the deadline, all I gotta do is edit this data into new sections and improve some of the old ones.

Two days ago I went to this Nomikai (work-related drinking “parties”) with some industry contacts of my lab. While the nomikai itself was not very exciting (I don’t really like this kind of Japanese event, but that’s for another post), I had a nice little neat idea while chatting there.

One of the research topics addressed in my laboratory is the use of Evolutionary Computation to assist in music composition. Basically, a EC algorithm generates multiple small pieces of music, which are evaluated by the human composer, and those evaluation scores are sent back to the computer, which try to generate a new generation of pieces similar to those which received a high score. This particular framework of evolutionary computation is called “Interactive Evolutionary Computation” (IEC) [1], because the fitness function is a human operator, and not a algorithmic function.

A big issue IEC is “user burden”. Evolutionary computation is based on scoring multiple candidate solutions, many times - when this evaluation is done by a human, instead of a computer program, the user may get tired after scoring too many individuals. To avoid that, it is important to either use the least amount of evaluations as possible, or make the evaluation as quick and painless to the user as possible - a lot of research has been done in both areas.

Now, the idea - how about using the concept of crowd sourcing to IEC? Instead of having one user evaluating the songs, we would have multiple users evaluating them in a asynchronous manner. The example we thought up would be a website where, say, mobile ring tones are generated by EC, with downloads and user evaluation being used as scores. Every few days(?), these values would be used to generate new tones, which would replace the old ones. This could not only generate more interesting tones, but also be able to “track” or “follow” fashions or memes of users.

A quick google search on the above keywords seemed to reveal that this is still a new idea (nothing relevant shows up on the first page for “crowd-sourcing IEC” and “crowd-sourcing composition” only show non-EC approaches [2]). Try it while it is fresh. Brainstorming in the comments is welcome

Links
[1] IEC on Wikipedia
[2] Crowdsourcing Composition

Today in the University meeting my professor told me that my pre-thesis defense will consist of an one hour presentation, followed by an one hour Q&A session. Since my Master Thesis presentation 2 years ago here in Tokyo university was a paltry 20 minutes, I was both relieved and a bit apprehensive. I was thinking that maybe one hour was a bit too much (I was expecting more like 40 minutes), but talking to Y I realized that among the three techniques and two problem I will have to explain and discuss at length, one hour might even be too little. Anyway, I’m breathing a little easier now that I know exactly how much time I have available - now I just need to do the work. I just wished they would give me the damn deadline already so I could prepare my schedule better.

Also, In today’s meeting we had visitors from another laboratory which presented to us their research on DNA computing. DNA computing is a sort of wet computing where you use the chemical reactions between DNA strands as the processing units. They were explaining their work in developing an AND gate with DNA. To be honest, I was not very impressed. I had heard before of wet computing before (maybe chemical computing?), and in my mind the state of art in this was a bit more evolved. But in their presentation, one AND operation would take more than one hour to complete, and they would need to do the experiment from scratch to change the data inputs. I wasn’t very convinced (although the rain might have made me grumpier than usual). Either you try to emulate electro-mechanical computing, but do it faster, or you get some new operators to do different stuff (like quantum computing). Could someone enlighten me about what I’m missing here?

In other news, RPG game tomorrow, and I got a pretty neat series of encounters for my players. Report coming from Sunday on

With a little less than 1 year to finish my PhD, the chief complaint from my advisor about my project is that, while my methodology has produced very good results, I have applied it to only one problem domain: Portfolio Optimization. While I have been making analysis and hybrids of my technique that have given me small differential gains in performance, what I really need is to find out other problem domains where I could apply my technique, to demonstrate its generality.

Unfortunately, it is not so easy as it seems. My MTGA, in short, was made for bounded parameter optimization problems. This means a problem that can be described as: “you have a value, X, which must be divided among Y variables (sum of all Y is X)”. The problem is that most benchmark functions for parameter optimization are not bounded. I have to assign the values of Y, but they are not bounded by an X sum. I tried to get around that by adding a “dump” variable which would get the remaining value of X, but that breaks down because I still need to know the max X, negative values make everything go wonky.

In my current research, X is the total resource available for investiment, and Y are the different assets that I can invest that resource into. If you can describe your research problem (in any field, no need to be remotely related to computing) in terms of this X and Y, I would love to hear about it!