Add configs/halfcheetah_6x1.yaml

configs/halfcheetah_6x1.yaml

@@ -0,0 +1,62 @@
+# HalfCheetah 6x1 — 6 agents, each controlling 1 joint
+# Used for main comparison (baseline vs PAL vs ours) and ablations
+
+environment:
+  env_name: mujoco
+  scenario: HalfCheetah-v2
+  agent_conf: "6x1"
+  agent_obsk: 1
+  episode_length: 1000
+  n_agents: 6
+
+training:
+  algorithm: mappo_lagrangian
+  num_episodes: 1500
+  n_rollout_threads: 10
+  num_env_steps: 15000000
+  share_policy: true
+  use_centralized_V: true
+
+network:
+  hidden_size: 256
+  layer_N: 2
+  activation: relu
+  use_feature_normalization: true
+  use_orthogonal: true
+  gain: 0.01
+
+optimizer:
+  lr: 5.0e-4
+  critic_lr: 5.0e-4
+  opti_eps: 1.0e-5
+  weight_decay: 0
+
+ppo:
+  clip_param: 0.2
+  ppo_epoch: 15
+  num_mini_batch: 1
+  entropy_coef: 0.01
+  value_loss_coef: 1.0
+  max_grad_norm: 10.0
+  use_gae: true
+  gae_lambda: 0.95
+  gamma: 0.99
+
+safety:
+  cost_limit: 1.0
+  lr_lambda: 1.0e-2
+  lambda_init: 1.0
+
+# Method-specific (ours only):
+responsibility_decomposition:
+  rho_mode: state_dependent
+  lr_rho: 5.0e-3
+  rho_update_period: 50  # T_rho: update rho every 50 episodes
+  cg_steps: 10           # conjugate gradient steps for implicit diff
+  rho_hidden_size: 64
+  rho_layers: 2
+
+seeds: [0, 1, 2, 3, 4]
+
+# Alpha weights (welfare importance per agent)
+alphas: [3.0, 2.5, 2.0, 1.5, 1.0, 0.5]