Hi, after reading through the control track in detail, we’re seeking several clarifications and also potentially found a bug in the code.
- What are the weights for the different reward metrics? The code disagrees with the AIcrowd description. E.g. the grid control weights (w3) are 0.075 each in the code but 0.125 on the website.
- The website has mentioned that the final evaluation will use different seeds for the power outages than Phase II online evaluation (is this correct?). However, will all the other environmental conditions (e.g. weather, buildings) be the same as Phase II online evaluation? We did not find a definitive answer on the website. Additionally, is the LSTM dynamics model for buildings 1/2/3 held constant across Phase II online/offline evaluation as well as final evaluation? In general, it would be helpful if all differences between Phase II offline, Phase II online, and final evaluation could be summarized.
- It is mentioned that Phase II online evaluation uses different weather and more buildings than Phase II offline evaluation, which only requires cooling but not heating. We’d just like to confirm that Phase II online evaluation and the final evaluation will also not require heating right?
- We’re fairly sure that there’s a bug in the battery charging code. Around lines 570-571, the
self.round_trip_efficiencymultiplier is erroneously applied in the more recent commit to both
self.energy_initas well as
energy, when it should only be applied to
energyalone (as in the previous version of the code). This could have implication for model-based methods?
Thanks for putting together the competition!