Agents.yaml

behaviors:
  AgentBehavior:
    trainer_type: poca
    summary_freq: 50000
    time_horizon: 256
    max_steps: 5e7
    keep_checkpoints: 5
    even_checkpoints: false
    checkpoint_interval: 500000
    threaded: false

    network_settings:
      hidden_units: 128
      num_layers: 2
      normalize: false
      vis_encode_type: simple
      conditioning_type: hyper

    hyperparameters:
      # Common Trainer Configurations
      learning_rate: 3e-4
      batch_size: 2048
      buffer_size: 131072
      learning_rate_schedule: linear

      # PPO/MA-POCA-specific Configurations
      beta: 5.0e-3
      epsilon: 0.2
      lambd: 0.99
      num_epoch: 3

    reward_signals:
      extrinsic:
        strength: 1.0
        gamma: 0.99

    self_play:
      save_steps: 100000
      swap_steps: 50000
      window: 30
      play_against_latest_model_ratio: 0.5
      initial_elo: 1200.0

checkpoint_settings:
  run_id: poca
  force: false

engine_settings:
  width: 84
  height: 84
  quality_level: 1
  time_scale: 1
  target_frame_rate: -1
  capture_frame_rate: 60
  no_graphics: true
  no_graphics_monitor: true

env_settings:
  num_envs: 1
  num_areas: 9
  seed: 42