One more

scripts/mig23-configs/circle40-nogoalnstuff/3-pureenergy-nodisc.yaml

@@ -0,0 +1,165 @@
+---
+trainer:
+  steps: 200
+  workers: 2
+
+  tensorboard_name: pureenergy
+  save_freq: 100
+
+  PPOConfig:
+    optimizer: adam
+    OptimizerKwargs:
+      lr: 0.0001
+      betas: !!python/tuple [0.9, 0.999]
+      eps: 1.0e-07
+      weight_decay: 0
+      amsgrad: false
+
+    gamma: 1.0
+    eta: 0.0
+    gae_lambda: 0.928
+
+    use_ugae: true
+
+    advantage_normalization: true
+
+    eps: 0.122
+    target_kl: 0.0445
+    entropy_coeff: 0.01387
+    entropy_decay_time: 100
+    min_entropy: 0.0
+    value_coeff: 1.0
+    rewind: true
+
+    ppo_epochs: 20
+    minibatch_size: 2048
+
+    use_gpu:
+
+model:
+  activation: leaky_relu
+  initializer: orthogonal
+
+  sigma0: 0.4
+
+  beta: false
+
+  hidden_sizes: [64, 64, 64]
+  separate_value: true
+
+
+  vec_hidden_layers: [512, 512, 512, 512]
+  rel_hidden_layers: [512, 512, 512, 512]
+  com_hidden_layers: [512, 512, 512]
+  emb_size: 64
+  attention_heads: 2
+
+
+environment:
+  num_agents: 40
+  initializer: "Circle"
+  dynamics: "PolarAcceleration"
+  observer: "Egocentric"
+  rewarder: "DecisionRewarder"
+
+  # Physics
+  max_speed: 4.0
+  max_acceleration: 5.0
+  rotation_speed: 3.0
+
+  # Spawn
+  spawn_noise_scale: 1.5
+  spawn_scale: 7
+
+  grid_spawn: true
+  group_spawn_scale: 1.5
+
+  enable_obstacles: true
+  block_scale: 3
+
+  random_mass: true
+  random_energy: true
+
+  shared_goal: true
+
+  # Unified reward
+
+  r_bmr: 0.0
+  r_drag: 0.0
+  r_dynamics: 0.0
+  r_potential: 0.0
+  r_diff_potential: 0.0
+  r_s_matching: 0.0
+  r_s_matching_exp: 0.0
+  r_speeding: 0.0
+  r_exp_v_matching: 0.0
+  r_exp_v_sigma: 0.85
+  r_v_matching: 0.0
+  r_goal: 0.0
+  r_final: 0.0
+  r_avg_final: 0.0
+
+  collision: -0.05
+
+
+  # Observations
+  sight_radius: 10
+  sight_agents: 20
+  sight_angle: 180
+  sight_acceleration: false
+
+  rays_per_direction: 10
+  ray_length: 10
+  ray_degrees: 90
+  ray_agent_vision: true
+  destroy_raycasts: true
+
+
+
+
+  evaluation_mode: false
+  # save_path:  # This should probably be set manually
+  early_finish: false
+  nice_colors: true
+
+  show_attention: false
+  backwards_allowed: false
+
+  # Legacy reward function
+  potential: 1
+  goal: 10
+  step_reward: -0.005
+
+  standstill_weight: 0
+  standstill_exponent: 0
+  goal_speed_threshold: 0
+
+  comfort_speed: 1.33
+  comfort_speed_weight: -0.75
+  comfort_speed_exponent: 1
+
+  comfort_distance: 0
+  comfort_distance_weight: 0
+
+  energy_weight: 1
+  final_energy_weight: 1
+  potential_energy_scale: 2
+  complex_energy: true
+
+env_curriculum:
+  1:
+    r_potential: 1.0
+    r_s_matching_exp: 2.0
+    r_speeding: 10.0
+    r_goal: 10.0
+  200:
+    # removed
+    r_speeding: 0.0
+    r_potential: 0.0
+    # same
+    r_goal: 0.0
+    # new
+    r_bmr: 1.0
+    r_dynamics: 1.0
+
+model_type: "relation"