Some questions I've received, with my answers.

11/11/05: How do you compare an algorithm that learns quickly but stabilizes at a lower level, vs one that learns slowly but gets to a better policy in the end?
We will present the entire curve. We will consider several ways of reducing these curves to single numbers to see how differently they rank the solvers. If you can tune your solver to maximize a particular quantity, I'd recommend maximizing the area under the learning curve.
11/11/05: Is there an English explanation of the sensor-network problem?
More details are now available on the main web page.
11/14/05: I do not think I will be able to visit the workshop at NIPS, is that a problem?
No, we will present whatever data is available. Those who attend will have the opportunity to answer questions about their system and discuss the results.
11/14/05: Can I use my own reward functions for the continuous control problems?
I would consider such a function as background information, which can be used but should be documented.
11/14/05: Can we use algorithms designed for continuous action spaces?
We have no domains with continuous actions in this event. Perhaps in future years.
11/14/05: Is it allowed to copy the dynamic plant functions from the benchmark environments? Or do I have to learn the plant dynamics explicitly with another function approximator?
The point of the event is to learn about the environments via interaction. However, if you supply the plant dynamics as background knowledge, you should document it.
11/14/05: Will you consider the used function approximator for the value function (and/or the policy) in your final rankings?
Unfortunately, we were only able to put together a small number of simple domains this year, so some powerful algorithms may be at a (relative) disadvantage Be sure to document how your results were generated and we will make this information available.
11/14/05: Shall I submit results for standard algorithms that I have implemented or are just the best algorithms useful for you?
We're interested in lots of different algorithms, so if you have multiple algorithms implemented, feel free to submit results from them. Remember that this year's event is not a competition, so finding a "winner" is not our the primary goal.