Hi AICrowd team, would like a couple clarifications:
resources must be carefully managed simultaneously in many buildings. In the CityLearn introduction paper, it was stated that the grid would impacted by interactions between different buildings. As pricing / CO2 cost is given in a CSV file (is this a constant value across different episodes?), how would the buildings impact each other wrt. the score of district CO2 and price?
evaluation criteria I am confused on exact evaluation method used. Do the 5/17 buildings have access to the data of the rest of the grid? Also, will the evaluation simulation take place using the same environment (year) data as the training phase?
Resource limitations: “The simulation period is 8,760 time steps i.e. one year, and participants can train on as many episodes of the simulation period, as needed.” What are the rules of pretraining an agent, and are there any resource / time constraints imposed on a submitted agent?
I’ll address your question regarding resource constraints.
Your agent need to complete 5 episodes in 60 minutes. I’ve also updated this in the stater kit documentation.
The resources it has access to is a machine with 4 vCPUs and 16 GB of RAM. The environment does not run on this machine so you should have full access to the resources.
As for training, yes, feel free to create an agent however you like. However, please note that the future phases will have private schemas that will be different from the current public schema, so it is advisable not to overfit your agents to the current schema.
Hi @mark_haoxiang. Thank you for your questions and I am happy to answer questions 1 and 2:
resources must be carefully managed simultaneously in many buildings - Yes, pricing and CO2 time series are constant across different episodes since each episode uses the same year’s data. The time series themselves vary through the year. The second part of the question is not clear to me but here’s my attempt at answering (please, let me know if this is clear and answers your question): Buildings do not impact each other, rather the manner in which electricity is stored in and discharged from the batteries will affect the electricity that is drawn from the grid hence, affecting the district CO2 and price.
evaluation criteria - No, the 5/17 buildings do not have access to the data of the rest of the grid. Also, in Phase I (training phase) of the competition, you are correct that the evaluation simulation uses the same year data as the training phase. By Phase II and Phase III of the competition, the remaining 12 buildings will be introduced in the evaluation simulation but will not be available for local evaluation where only the 5/17 buildings from the Phase I are provided.
Hope these answer your questions otherwise, I am happy to clarify further.
You say that CO2 time series are constant across different episodes but also that buildings affect the district CO2 emissions. While I understand the second statement, I don’t understand the first one.
In Competion Overview I read that “the challenge utilizes 1 year of operational electricity demand and PV generation data”; this seems reasonable since they are uncontrollable variable (they depends on users and local weather…) while the global CO2 emission of the district seems to be right the variable that the agents must collectively optimize. Looking at the possible states of the environment, CO2 even looks like the only variable (carbon intensity) that may depend on collective behaviour; I don’t see variables related to the energy price for example. Are these deductions right?
Sorry if the question is too long but I’m currently studying the environment and I’d like to better understand how it works!
Hi @alberto_ingenito. Thanks for your questions. The first statement that CO2 time series (kg_{CO_2}/kWh) are constant just meant that the same file is used in all episodes. Sorry for the confusion.
For your second question abut the states, you are looking at a different branch, master which, isn’t being used for the CityLearn Challenge 2022. CityLearn Challenge 2022, please refer to the citylearn_2022 branch. We have not merged both branches yet. The energy price variables i.e. states are defined in the schema.json, and come from the pricing.csv file.