Hi, and thank you for releasing LC-Opt — the FMU, the Gym env, and the pretrained CA/Multihead-CA checkpoints are a great resource.
We're reproducing the Table-3 benchmark on the shipped LC_Frontier_5Cabinet_4_17_25.fmu and ran into a few reproducibility questions about the exact evaluation setup. Would appreciate any clarification.
1. The demanding "unseen exogenous trace" used for Table 3 / Table 4.
On the shipped input_04-07-24.csv, the ASHRAE-G36 baseline holds the blade-temperature band [20, 40]°C at ~100% (boundary temps sit ~26–36°C). Both the Gym env and frontier_env.py's exogenous_variable_generator clip every cabinet load column to [mean − 1.75σ, mean + 0.1σ] and then divide by (cabinets × branches), which keeps blade temps well inside the band — so G36 never approaches the 76.92% D_blade reported in Table 3. Could you release (or describe) the exact exogenous trace used for the Table-3/Table-4 evaluation? Is it a held-out segment of input_04-07-24.csv, a scaled variant, or a separate file that isn't in the repo?
2. Cooling-tower configuration behind the ~237 kW / 206 kW numbers.
Table 3 reports CT power ≈ 237.31 kW (G36) and 206.52 kW (SOTA). The shipped LC_Frontier_5Cabinet FMU exposes only two CT cells (...cell[1].CT.PFan, ...cell[2].CT.PFan), each with PFan_nominal ≈ 82 kW → a ~164 kW ceiling. Was Table 3 generated on a larger configuration (more CT cells / cabinets, e.g. one of the Table-4 scaled configs)? If so, which FMU / config produces the Table-3 numbers?
3. D_blade definition.
Just to confirm: D_blade is the percentage of timesteps for which all blade-group boundary temperatures (...cabinet[1].boundary_{1,2,3}.port.T) lie within [20, 40]°C (per §6), averaged over the evaluation episode — correct? And the same blade-temperature variable set is used across all cabinets?
Context: we'd like to compare against your reported numbers fairly. Without the exact eval trace + CT config, we can only reconstruct a demanding regime ourselves, which isn't apples-to-apples with Table 3. Any pointers (or a data drop) would let the community reproduce the headline result directly. Thanks!
Hi, and thank you for releasing LC-Opt — the FMU, the Gym env, and the pretrained CA/Multihead-CA checkpoints are a great resource.
We're reproducing the Table-3 benchmark on the shipped
LC_Frontier_5Cabinet_4_17_25.fmuand ran into a few reproducibility questions about the exact evaluation setup. Would appreciate any clarification.1. The demanding "unseen exogenous trace" used for Table 3 / Table 4.
On the shipped
input_04-07-24.csv, the ASHRAE-G36 baseline holds the blade-temperature band [20, 40]°C at ~100% (boundary temps sit ~26–36°C). Both the Gym env andfrontier_env.py'sexogenous_variable_generatorclip every cabinet load column to[mean − 1.75σ, mean + 0.1σ]and then divide by (cabinets × branches), which keeps blade temps well inside the band — so G36 never approaches the 76.92%D_bladereported in Table 3. Could you release (or describe) the exact exogenous trace used for the Table-3/Table-4 evaluation? Is it a held-out segment ofinput_04-07-24.csv, a scaled variant, or a separate file that isn't in the repo?2. Cooling-tower configuration behind the ~237 kW / 206 kW numbers.
Table 3 reports CT power ≈ 237.31 kW (G36) and 206.52 kW (SOTA). The shipped
LC_Frontier_5CabinetFMU exposes only two CT cells (...cell[1].CT.PFan,...cell[2].CT.PFan), each withPFan_nominal ≈ 82 kW→ a ~164 kW ceiling. Was Table 3 generated on a larger configuration (more CT cells / cabinets, e.g. one of the Table-4 scaled configs)? If so, which FMU / config produces the Table-3 numbers?3.
D_bladedefinition.Just to confirm:
D_bladeis the percentage of timesteps for which all blade-group boundary temperatures (...cabinet[1].boundary_{1,2,3}.port.T) lie within [20, 40]°C (per §6), averaged over the evaluation episode — correct? And the same blade-temperature variable set is used across all cabinets?Context: we'd like to compare against your reported numbers fairly. Without the exact eval trace + CT config, we can only reconstruct a demanding regime ourselves, which isn't apples-to-apples with Table 3. Any pointers (or a data drop) would let the community reproduce the headline result directly. Thanks!