• Crosery's avatar
    v1: simplified MLP+GRU model with improved rewards · b1238039
    Crosery authored
    Features: 67D vector + 5ch 21x21 map (2272D total)
    Model: Entity MLP + CNN + State MLP + Fusion + GRU, single value head (~172K params)
    Reward: log monster shaping, balanced dense signals, curriculum learning 5 phases
    GRU hidden: ObsData→model→ActData→update_status flow (参照参赛代码LSTM模式)
    b1238039
This project is licensed under the Other. Learn more
license.dat 12 KB