PPSN 2008, post 3

With some hindsight - that is, one day's worth of hindsight - I must say that PPSN 2008 was one of the best conferences I've ever attended. On my very personal ranking, it's up there in the top with CIG 2007 in Hawaii. The organisation of PPSN was top-notch, with nothing going wrong, the food good, the opportunities for socialising/networking plentiful and the schedule adhered to (this is Germany, after all).

It's striking how high the quality of the papers are in PPSN. In Gecco and CEC you know and then find papers you think should not have been accepted in a scientific conference, but never so in PPSN. There is also a difference in "scientificness". Many papers at the two other major conferences present yet another variation on a well-known algorithm, or yet another application, with little in the way of analysis and comparison to the state of the art. At PPSN, the norm seems to be to isolate a particular phenomenon, parameter or operator of known algorithms and benchmarks and study it further, making sure that even papers that are not groundbreaking (which is by necessity most papers) add to the body of human knowledge.

Of course, there are new algorithms and applications at PPSN as well. In particular, a number of variants of the CMA-ES were presented. CMA-ES seems to have become the standard algorithm to benchmark continuous optimization algorithms against, which makes sense, as it reaches good result very quickly on many problems.

Speaking of benchmarks, there seems to be a consensus that there is a lack of good reinforcement learning benchmark. A poster by Marc Schoenauer even went so far as to list "Stop balancing the double pole!" among it's conclusion. Of course, I tried to convince everybody who brought up the topic that they should use simplerace instead. A much better benchmark in many ways.

Now I'm going to prepare the talk I'll give tomorrow (with Marie Gustafsson) on "AI from Science to Fiction" at the Fantastic Film Festival in Lund, and the talk I will give on Monday on "Computational Intelligence and Game Design" at ITU Copenhagen.

PPSN 2008, post 2

We've now been to the conference dinner, in an old steel mill. Excellent food, and an excellent place to visit. I love lo-tech and especially huge, rusting metal structures.

Towards the end of the conference dinner, DJ JJ organized a singing competition between Spanish-speakers and a group of speakers of various other tongues(Swedish, English, Dutch, Hebrew etc.). Needless to say, we lost - we couldn't find any songs in English we all knew the lyrics to. After the dinner most of us went back to our respective hotels to recuperate. In my case, I'm recuperating from the excellent Dortmund nightlife. Clubbing until 4.30 on a packed dance floor, with the beers priced 50 cents each. On a Monday. Brilliant!

So you want to know about the scientific side of things? Well, PPSN is really a very good conference. It is hard to find a paper which is not good, though it is of course easy to find papers that don't interest me. I'm not interesting in everything. For example, a theoretical analysis of the behaviour of the 1+1 ES does not interest me very much, even if it is obvious that the research is sound and the paper good. Still, there are many papers here that I like.

PPSN 2008, post 1

I am at PPSN in Dortmund. I am listening to the talk on "Semidefinite Programming and Lift-and-Project Methods in Combinatorial Optimization" by Levent Tunçel. I do not understand anything at all of this talk. I am not the only one in the audience.

While I applaud the PPSN tradition of inviting people who are not evolutionary computation researchers to give keynotes, I don't think people outside our field understand how very little mathematics many (most?) people in this field know.

But the tutorials yesterday and the workshops on Saturday were really nice.

What should we call non-evolutionary reinforcement learning?

I work in evolutionary reinforcement learning. That is, I develop reinforcement learning problems, and evolutionary algorithms that solve such problems.

The problem is that many people in reinforcement learning (RL) would say that I'm not working on RL at all. These people work on things like temporal difference learning, policy gradients, or (more likely) some newfangled algorithms I have never heard of, but which are most certainly not evolutionary. Most of the people working on non-evolutionary RL probably don't know much (maybe nothing!) about evolutionary RL either. So disconnected are our communities. It's a shame.

In their discipline-defining (4880 citations on Google Scholar) book "Reinforcement Learning", Sutton and Barto start with defining RL as the study of algorithms that solve RL problems, and mention in passing that they can be solved by evolutionary algorithms as well. The book then mentions nothing more about evolution, and goes on to essentially discuss TD-learning and variations thereof for a few hundred pages.

In practice, the evolutionary and non-evolutionary RL folks publish in different conferences and journals, and don't cite (nor read?) each other much. We write our papers in very different styles (the non-evolutionary RL people having much more maths in them, evolutionary RL researchers often relying on qualitative argument coupled with experimental results), and I for one often simply don't understand non-evolutionary RL papers.

Again, it's a shame. And it would be great if we could find some way of bridging this divide, as we work on the same class of problems.

But to do this, we need to find a way of addressing the issue, which was really the purpose of this blog post. Simply put, what do we call the two classes of algorithms and the research communities studying them? This is an issue I run into now and then, most recently when writing a grant proposal, and now again when preparing lecture slides for a course I'll be teaching this autumn.

The non-evolutionary RL people would not want a negative definition, based on what their algorithms aren't rather than what they are. They would rather go for RL, plain and simple, but this has the problem that non-evolutionary RL is excluded from that field, in spite of being part of the definition. In the proposal we wrote we ended up talking about "classical" versus evolutionary RL, but this has the problem that evolutionary algorithms predated td-learning by several decades. We could also use the term "single-agent RL", but then again, a simple hill-climber is arguably a (degenerate) evolutionary algorithm, and very much single-agent. Besides, there is multi-agent non-evolutionary RL. Sigh.

So I really don't know.

CIG 2008 special session on coevolution in games

Coevolution in Games: A Special Session at IEEE Symposium on Computational Intelligence and Games
Perth, Australia
15-18 December, 2008

Special Session Chairs: Julian Togelius, Alan Blair and Philip
Hingston
Contact: julian@togelius.com

Submission deadline: 15 August 2008

http://www.csse.uwa.edu.au/cig08/specialSessions.html

Description:

In coevolution, the fitness of a solution is determined not (only) by a fixed fitness function, but also by the other solution(s) being evaluated. Thus, coevolution has the potential to overcome several problems with static fitness functions, paving the way for more open-ended evolution. However, several phenomena common to coevolutionary algorithms are at present poorly understood, including cycling and loss of gradient. Further understanding of such phenomena would facilitate more widespread use of coevolutionary algorithms.

This special session seeks to bring together research that uses coevolutionary algorithms to learn to play games, uses games to investigate coevolution, or uses coevolution as a basis for game design. Due to their adversarial nature, often involving interaction of multiple agents, games are uniquely suited to be combined with
coevolution. We invite both theoretical and applied work in the intersection of coevolution and games, including but not limited to the following topics:

Competitive coevolution
Cooperative coevolution
Multiple populations in coevolution
Coevolution with diverse representations
Theory of coevolution
Preventing cycling and loss of gradient
Coevolution-based game design
Self-play and coevolutionary-like reinforcement learning
Relative versus absolute fitness metrics

About the organisers:

Julian Togelius is a researcher at the Dalle Molle Institute for Artificial Intelligence (IDSIA) in Lugano, Switzerland. His research interests include evolving game-playing agents, modelling player behaviour, and evolving interesting game content, mainly using evolutionary and coevolutionary techniques. He also co-organizes the well-attended Simulated Car Racing Competitions for the IEEE CIG and CEC conferences. See his home page for more information.

Philip Hingston is an associate professor of computer science at Edith Cowan University in Perth. His research interests are in the theory and application of artificial intelligence and computational intelligence. He has a particular interest in evolutionary computation as a tool for design, and in computer games. He is chair of the IEEE CIS Task Force on co-evolution. More information can be found on his home page.

Alan Blair is Chair of the IEEE CIS Task Force on Co-evolution and Games. His research interests include robot navigation, image and language processing as well as co-evolutionary learning for Backgammon, Tron, IPD, simulated hockey and language games. Homepage.

Busy days in Hong Kong in June

If you're going to WCCI2008 in Hong Kong in June, you might run into me. Chances are, however, that I'll just keep on running, as I've got a busier schedule than I've ever had in a conference. So just don't take it personally, OK?

To begin with I'm giving a tutorial together with Simon Lucas and Tom Runarsson on "Learning to play games". Preliminarily, Tom will talk about evolution versus other types of reinforcement learning, Simon will talk about different sorts of function approximators, including his n-tuple classifier, and I will discuss different ways in which such techniques can be used in games, with examples. Not only car racing, is the plan...

Then, I will participate in a panel discussion about Computational Intelligence and Games, together with some of the people that are defining this young field.

Staying on the topic of games, I will then present the results of the car racing competition, together with Daniele Loiacono and Pier Luca Lanzi. By the way, have you considered participating? We now have proper software packages available, so it's easy to get started with developing controllers using your favourite brand of learning algorithms.

Yes, I do have some papers as well. However, these are not (primarily) about games this time...

The first paper, a collaboration with my good friends Alberto Moraglio and Renzo De Nardi, is called Geometric PSO + GP = Particle Swarm Programming. The idea is to combine Alberto's geometric particle swarm optimization algorithm with genetic programming. The good thing with GPSO is that it can be used in any search space where a distance between two solutions can be defined, and a weighted recombination operator for these solutions can be devised; in contrast, standard PSO only works in continuous spaces. We have previously successfully used GPSO to solve Sudoku problems, to show this really works. In our new paper, we invent a few weighted recombination operators for expression trees, and test the PSP (GPSO + GP) algorithm on two standard GP benchmarks. The results are decent, but not as good as we would have wanted them; we think this is down to the recombination operators still needing some more work. Nevertheless, we think this is the first time PSO has been applied directly to GP.

The other paper, my first collaboration with Faustino Gomez and Juergen Schmidhuber (my new colleagues at IDSIA) is called Learning what to ignore: Memetic climbing in topology and weight space. It's about "memetic neuroevolution", an idea I've had for a while, and which someone else really should have thought about. The key algorithm here (the "memetic climber") is incredibly simple: evolve topology of weights of a neural network at different time scales. Topology mutations are almost always very destructive, so after each topology mutation, do local search (hill-climbing) in the space of neural network weights do find the potential of the new topology. If the fitness at the end of the local search is not at least as high as it was before the topology mutation, the topology mutation is discarded. This algorithm turns out to work really well, especially for problems where an agent has to handle high-dimensional input arrays. In such problems, which is really the ones i'm most interested in, most neuroevolution algorithms get stuck in local optima if you don't manually specify a good topology for the neural net. So I have big plans for this little algorithm in the future.

Mentioned at the GDC

A paper by me, Renzo and Simon was recently mentioned as "one of the top ten research finding in games studies" at the Game Developers Conference. This is really cool, as GDC is the most important conference for the games industry, attended by all the important developers and industry heads and tons of journalists.

The presentation can be found here:
http://www.avantgame.com/top10.htm

A short description and some discussion on Raph Koster's site:
http://www.raphkoster.com/2008/02/20/gdc2008-game-studies-download-30/

And the paper:
http://julian.togelius.com/Togelius2007Towards

Quite surprising to see that I'm doing "games studies", I thought I was doing artificial intelligence...

Call for participation: IEEE WCCI 2008 Simulated Car Racing Competition

*** SIMULATED CAR RACING COMPETITION @ IEEE WCCI 2008 ***
*** DEVELOPING CONTROLLERS FOR A REALISTIC CAR RACING SIMULATOR ***

Homepage: http://cig.dei.polimi.it/?page_id=5

Deadline: May 25th, 2008
Organizers: Daniele Loiacono, Julian Togelius and Pier Luca Lanzi.
Email: carracing@gmail.com

The car racing competition, organized in association with the 2008 IEEE World Congress on Computational Intelligence WCCI 2008, is now open for submissions. Competition rules and further details can be found on the competition web site:

http://cig.dei.polimi.it/?page_id=5

while recent news can be found at

http://cig.dei.polimi.it/?cat=4

*** GOAL & SCOPE***

The goal of the competition is to develop racing car controllers that can compete on a set of tracks against the timer or other controllers.

There are no limitations on the type of controller, handcoded controllers are also welcome, although the focus of the competition, given the venue, will be on controllers using computational intelligence techniques (e.g. genetic programming, neuroevolution, td-learning, policy gradient search, fuzzy logic, etc.).

*** ORGANIZATION ***

The competition is organized as the successful car racing competitions held in association with CEC2007 and CIG2007, but it is based on "The Open Racing Car Simulator" (TORCS), a realistic car racing simulator, and structured as a client-server architecture.

A racing server will manage the race, while controllers will run as different processes. Communication between the racing server and the controllers is done through TCP/IP. The race is therefore in *real time* and the controller's reaction time is likely to be an important factor in the design.

We will score every submitted controller on the distance raced in a fixed amount of time when driving on its own on a set of tracks. At the end of the competition, the best few controllers will race against each other on a different set of tracks, validating that the controllers perform well in the presence of other cars and that their performance generalizes to other tracks than those they were trained for. The winner of the final competitive races will get to present their controller at WCCI2008 <http://www.wcci2008.org>, and will have their registration fee reimbursed.

*** SOFTWARE ***

The competition software can be downloaded from:

http://cig.dei.polimi.it/?page_id=5

where detailed installation instructions, a description of the sensors model, and examples of controllers are available; more example trainers and interfaces in different programming languages will be added soon.

On industrial-academic collaboration in game AI

Today was the kick-off event day for the UK Artificial Intelligence and Games Research Network. I wasn't there, but my ex-supervisor is one of the organizers, so I heard about it beforehand.

This is the first blog post I've seen about the kick-off; the author seems to have left the event with a pretty pessimistic view of the prospects for industrial/academic collaboration. His main complaints is that academics don't understand games, and the specific needs of game developers. Well, then tell us! I would love to hear about specific problems in game development where evolution or some other form of machine learning or computational intelligence could matter.

Alex J. Champandard, in a comment on the same blog post, develops the point further. He asks:

So why do you need government funding for [applied games research]? It's a bit like admitting failure :-)

On the other hand, if [academics are] doing research for the sake of research, why do they need input from industry?

These questions can be asked for just about any research project in the interface between academia and industry. And yet companies happily keep funding PhD students postdocs, and even professors in a huge number of research fields, from medicinal chemistry to embedded systems design to bioinformatics. In some cases these collaborations/funding arrangements definitely seem strange, but apparently it makes economic sense to the companies involved.

I once asked an oil company executive (at a party! Now, stop bothering me about what sort of parties I go to...) why his company funds a professor of geology. His answer was roughly that it was good to have expert knowledge accessible somewhere close to you, so you know who to ask whenever you need to. Plus, a professor's salary wasn't really that much money in the grand scheme.

Now, game companies and oil companies are obviously very different sorts of creatures. I think the main opportunity for game companies would be to outsource some of their more speculative research - things that might that not be implementable any time in the near future, either because the computational power is not there yet, or because the technique in question would need to be perfected for a couple of years before deployment. Having a PhD student do this would be much more cost-efficient than assigning a regular employee to do it (especially with government funding, but probably also without), and frees up the employee for actual game development. In addition, the company's own developers might very well be too stuck in the way things currently work to try radically new ideas (of course, academics might also be stuck in old ways of thinking, but there are many academics around and if you offer some funding you can typically select which academic you want to work for you).

This argument assumes that game companies do any sort of research into technologies that lie more than one release cycle away. I'm not stupid enough to claim that no game companies do this - e.g. Nintendo obviously does - but I venture to guess there are many that don't.

As for the other part of Alex's question, "if we do research for the sake of research, why do we need input from industry?", the answer is more obvious. Because even if we do research because we love the subject itself and really want to find out e.g. how to best generalize from sparse reinforcements, we also want to work on something that matters! And fancy new algorithms look best together with relevant problems. It's that simple.

New webpage at IDSIA

I've now set up a parallel home page at my new workplace, IDSIA. It's currently mostly a quick overview of the various things I'm involved in academically, but I plan to set up pages with a bit more detail about the projects in that domain as well (for those who don't feel like going straight for the papers).

My primary home page will still contain my publication list, CV and such formal stuff.