Lux AI Challenge Copyright 2021 @kuto_bopro Meta Kaggle Collection of episodes ・Team: Toad Brigade ・LB score > 1900 ・only win game ・about 1000 episodes(3 submissions) Unet Imitation Learning approach inspired by nosound(@zharch) obs horizontal flip vertical flip random roll(-5~5) TTA obs obs Global features (8ch,4,4) Observation map (17ch, 32, 32) Policy map (3ch,32,32) ・Units counts (×2) ・Citytiles counts (×2) ・Research points (×2) ・turn / cycle Data Sampling ・Random sampling up to 4 units actions in each turn ・Downsampling center actions Extract units policy from each units position Image reference: https://www.lux-ai.org/ ・Units position/cooldown/resource (×2) ・Citytiles position/cooldown/fuel-lightupkeep ratio (×2) ・Wood/Coal/Uranium positions ・Road level ・Effective map area Create 8 pattern policy maps and apply mean UNet model Decide citytile actions by simple rule Create 4 batch by rotation input (4 batches) Policy maps (4batch, 3ch, 32, 32) Final policy map (6ch, 32, 32) Hierarchize move actions (shared by nosound) output 3ch policy map (4 batches) 90° 180° 270° 90° 180° 270° 0ch: Center Action → batch mean 1ch: Move Action 1st batch: north 2nd batch: west 3rd batch: south 4th batch: east 2ch: Build City Action → batch mean kuto(@kuto0633) Final policy map (6ch,32,32) Observation maps(4batch, 17ch, 32,32) 0ch: Move Center 1ch: Move North 2ch: Move West 3ch: Move South 4ch: Move East 5ch: Build City Calculate 4 move actions as one direction State Value (for RL and MCTS but not work) 16 64 64 128 128 256 256 256 256 8 256 256 +8 256 256 128 + 256 128 128 64 + 128 64 64 3 32×32 32×32 32×32 16×16 16×16 16×16 8×8 8×8 8×8 4×4 4×4 4×4 32×32 32×32 32×32 16×16 16×16 8×8 8×8 FC BN ReLU FC 264→64 64→1 Conv2d BatchNorm2d, ReLU MaxPooling2d Upsample Concatenate Private LB: 34th (score 1570)