# Default configuration for Wayfinder agent training

agent:
  max_actions: 1000
  latent_dim: 256
  buffer_size: 200000
  device: cpu  # or cuda

perception:
  latent_dim: 256
  conv_channels: [32, 64, 128]
  conv_kernel: 3

world_model:
  latent_dim: 256
  action_embed_dim: 32
  hidden_dim: 256
  lr: 0.0001
  buffer_size: 200000
  batch_size: 64
  entropy_reg: 0.01
  fwd_loss_weight: 0.5

intrinsic_reward:
  w_extrinsic: 1.0
  w_novelty: 0.3
  w_curiosity: 0.2
  novelty_decay: 0.95
  curiosity_clip: 10.0
  score_baseline: 0.0

planner:
  max_depth: 5
  max_simulations: 50
  exploration_constant: 1.414
  discount: 0.95

action_head:
  temperature: 0.5
  epsilon: 0.15
  coord_head_channels: [32, 16]

training:
  mode: online  # or offline
  epochs: 50
  batch_size: 64
  lr: 0.0001
  eval_interval: 10
  save_interval: 20

eval:
  games: ["ls20", "ls21", "ls22"]
  max_actions_per_level: 1000
  output_dir: "eval/results"