Add color_shared + attention 1000-iteration experimental config

- hidden_size: 256 (smaller for faster iteration) - num_layers: 2 - color_attention_layers: 2, color_attention_heads: 8 - max_iterations: 1000 - advantage/strategy_updates: 256 (reduced from 512) - Ready for experimental training run Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>
2026-05-07 09:21:50 +09:00
parent b7fce79f5b
commit cf67bc2198
1 changed files with 112 additions and 0 deletions
@@ -0,0 +1,112 @@
+run:
+  experiment_name: lost_cities_deep_cfr_color_shared_attention_1000iter
+  iterations: null
+  seed: 42
+  max_iterations: 1000
+  max_hours: null
+  device: cuda
+  use_amp: false
+
+rules:
+  n_colors: 5
+  n_ranks: 9
+  min_rank: 2
+  n_handshakes: 3
+  hand_size: 8
+  expedition_penalty: -20
+  bonus_threshold: 8
+  bonus_amount: 20
+
+encoding:
+  derived_playability: true
+  slot_aware_playability: true
+
+network:
+  kind: color_shared
+  hidden_size: 256
+  num_layers: 2
+  activation: relu
+  color_attention_layers: 2
+  color_attention_heads: 8
+
+traversal:
+  traversals_per_iteration: 2
+  traversals_per_player: 70
+  max_depth: null
+  max_nodes: 10000
+  max_nodes_per_traversal: 1000
+  regret_matching_epsilon: 0.0001
+  outcome_sampling_epsilon: 0.2
+  outcome_sampling_value_clip: 500.0
+  outcome_unsampled_regret: zero
+  cutoff_value_mode: score_diff
+  cutoff_rollouts: 0
+  cutoff_rollout_policy: random
+  cutoff_rollout_max_steps: 300
+  opponent_policy: self_play_league
+  strategy_sample_interval: 1
+  store_strategy_on_traverser_nodes: true
+  store_strategy_on_opponent_nodes: false
+  num_workers: 8
+  worker_chunk_size: 4
+  traversal_worker_chunk_size: 8
+  progress_every_traversals: 10
+  endpoint_depth_bucket_width: 100
+  endpoint_depth_bucket_max: 1000
+
+regret_matching:
+  all_negative_fallback: argmax_tiebreak
+
+training_weighting:
+  mode: none
+
+self_play:
+  snapshot_every: 1
+  max_snapshots: 20
+  anchor_probability: 0.0
+  current_weight: 0.45
+  recent_weight: 0.30
+  older_weight: 0.15
+  anchor_weight: 0.10
+  recent_window: 5
+
+optimization:
+  advantage_train_steps: 1
+  strategy_train_steps: 1
+  batch_size: 32
+  advantage_batch_size: 1024
+  strategy_batch_size: 1024
+  advantage_updates_per_iteration: 256
+  strategy_updates_per_iteration: 256
+  learning_rate: 0.00003
+  weight_decay: 0.0001
+  grad_clip: 1.0
+
+memory:
+  advantage_capacity: 2000000
+  strategy_capacity: 2000000
+
+checkpoint:
+  directory: runs/deep_cfr/deep_cfr_color_shared_attention_exp_1000iter
+  save_latest: true
+  save_every_iteration: false
+  save_iteration_interval: 100
+  save_latest_only: false
+  progress_interval_seconds: 20.0
+  exact_resume: false
+
+evaluation:
+  eval_every: 5
+  games: 100
+  opponents:
+    - random
+    - passive_discard
+    - safe_heuristic
+    - safe_heuristic_loose
+    - safe_heuristic_strict
+    - noisy_safe
+  max_steps: 10000
+  on_max_steps: score_diff
+  batch_size: 64
+  device: trainer
+  num_workers: 4