About the Optimizing Demand-side Strategies competition :
I have a concern about the test phase, will we be provided with the same site_id (the 11 from the train/submit data) or some brand new site_id? In the former, we should learn site features, in the latter, we should learn to generalize.
Or maybe we do not know what site_id there is in the test phase (and we should learn both)?
Moreover, in the first case, may we assume that price_buy and price_sell will behave (roughly) the same ways?
Thank you for your clarification
I have the same concern, I don’t understand why no one has already answer you.
Great question: Evaluation will be done using the same site ids as the available training data, and it is acceptable to build site-specific models provided that:
- The corresponding model training is automated and can be replicated for new sites.
- Only past data has been used to build the model that is used for predictions at time t.
Thanks for your clear answer bull. I have also a question regarding the number of simulations that will be evaluated at the end of the competition.
There is a constraint on the execution time: 30 minutes.
And it is also said that the submission will be evaluated against 10 day periods.
However the number of simulations is unknown and that is important to verify if the time constraint is satisfied by the solution.
We’re still finalizing exactly how many simulations, but you can use the provided “submit data” as the upper-bound for the number of simulations. Currently that file takes ~10 minutes to execute with no battery controller.
Will update when we have more clarity.
Only past data has been used to build the model that is used for predictions at time t.
Does it mean that we need to make models for each timestamp to make sure not to use future data?
Wouldn’t it be easier to test it against the data collected/simulated after the public ones?
Answered here in the problem description:
But that is a different challenge.
Can you confirm that the simulation with no battery takes 10 minutes on the Xeon 8175?
I’m trying to find information about the processor for comparing with mine but it seems that it is a custom processor made for AWS, so there is not too much information.
Yes, that is more or less right.
That said, we are planning to release additional information about how many simulations we will use for the final evaluation. As of right now, we expect that competitors will have additional time per simulation as compared to what we currently ask.
My current solution takes 27 minutes on my computer. It has an i7 processor, so I’m worried if those 3 extra minutes will be enough on the evaluation processor.
We’ve finalized the evaluation set–you shouldn’t have to worry. Here’s the annoncement:
Thank you very much @bull for this information.
I would just like to express strong worries about the choice of only 34 simulation periods for the evaluation.
Let’s just 1 minute consider the score of each participant as a random variable with some expected value and a standard deviation on a given simulation. Since the standard deviation scales as 1 / sqrt(N) for N independent simulations, the choice of 34 simulations only leads to a reduction of 6 of the variance on individual simulation. This is extremely low !
A number of 240 simulations, such as the submit, would already be high variance. But with only 34 simulations, you will not select the best solution but rather the one which had the best luck.
It is even more regrettable that I imagine your strong attention to this selection was due to fairness concerns. In reality, such a restricted selection of evaluation data leads to the opposite.
I ask you to please consider this matter seriously and add more data to the evaluation, so that the final score truly reveals the best solution and not random flukes. Some participants have invested a lot of time and would like to be evaluated fairly with sufficient data.
I think the same, changing the rules at one week to the end of the challenge is not fair. I had put a lot of effort in optimizing the execution time of my solution to fit the requirement time.
And suddenly time is not an important constraint. So I may have lost my competitive advantage agains someone who instead of optimizing the time has optimized the score.
Thanks for your thoughts, as always we try to be fair to competitors and to change things as infrequently as possible during the competition.
As part of our evaluation, we will ensure that the winning solutions are consistently the best performing–if necessary, by running additional simulations. This will be done at the sole discretion of the judges, but with an eye to fairness and ensuring that there is not a winner by chance.
Thanks again for participating and for your work and thoughtful questions throughout!