# SUMO + TraCI VSL training configuration

sumo:
  net_file: "sumo_resource/modified.net.xml"
  # route_file: "sumo_resource/routes_filtered.xml"
  route_file: "sumo_resource/routes_20231002_optimized.xml"
  # route_file: "sumo_resource/routes_optimized.xml"
  detector_add_file: "sumo_resource/metrics_il.add.xml"
  enex_add_file: "sumo_resource/metrics_enex.add.xml"
  step_length: 1
  begin_time: 54000
  end_time: 57600
  # begin_time: 39600
  # end_time: 57600
  # gui: true
  gui: false
  no_warnings: true

environment:
  control_interval: 60
  control_segment_length_m: 1000
  detector_spacing_m: 100
  passive_prefix_segment_count: 3

  # Ordered reference corridor edges. The environment derives 1000 m
  # control segments from this junction-bounded corridor.
  control_edges:
    - "G1523_AM3_4.1"
    - "G1523_AM3_4.2"
    - "G1523_AM3_4.3"
    - "G1523_AM3_4.4"
    - "G1523_AM3_4.5"
    - "G1523_AM3_4.6"
    - "G1523_AM4"
    - "G1523_AM5"
    - "G1523_AM6"
    - "G1523_AM6.1"
    - "G1523_AM6.2"
    - "G1523_AM6.3"
    - "G1523_AM6.4"
    - "G1523_AM6.5"
    - "G1523_AM6.6"
    - "G1523_AM7"
    - "G1523_AM7.1"
    - "G1523_AM7.2"
    - "G1523_AM8"
    - "G1523_AM8.1"
    - "G1523_AM8.2"

  # Available VSL actions in km/h
  speed_actions_kmh: [40, 60, 80, 100, 110]
  free_flow_speed: 30.56

  reward:
    # Positive term: throughput
    w_flow: 0.4
    # Negative term: speed dispersion across edges
    w_var: 0.3
    # Density-adaptive hard-brake penalty weight
    w_brake_base: 0.1
    w_brake_max: 0.3
    # Penalty for abrupt VSL changes
    w_penalty: 0.2

    rho_critical: 44.75
    k_sigmoid: 0.2
    d_threshold: 5.0
    d_max: 20.0
    C_max: 4924
    v_limit: 30.56
    delta_vsl_max: 16.67

training:
  num_episodes: 4000
  random_seed: 42
  log_freq: 10
  save_freq: 20

agents:
  common:
    device: "cuda"

  dqn:
    hidden_dim: 256
    learning_rate: 0.0003
    gamma: 0.99
    epsilon_start: 1.0
    epsilon_end: 0.01
    epsilon_decay: 600
    buffer_size: 20000
    batch_size: 128
    target_update: 5

  ppo:
    hidden_layers: [256, 256]
    learning_rate: 0.0003
    gamma: 0.99
    gae_lambda: 0.95
    clip_epsilon: 0.2
    value_coef: 0.5
    entropy_coef: 0.01
    max_grad_norm: 0.5
    ppo_epochs: 4
    batch_size: 15
    lr_schedule: "cosine"

  gpro:
    hidden_layers: [256, 256]
    learning_rate: 0.0003
    gamma: 0.99
    gae_lambda: 0.95
    clip_epsilon: 0.2
    value_coef: 0.5
    entropy_coef: 0.01
    max_grad_norm: 0.5
    ppo_epochs: 4
    batch_size: 15
    lr_schedule: "cosine"
    group_size: 4
    group_advantage_coef: 0.35
    advantage_epsilon: 1.0e-8

  appo:
    hidden_dim: 128
    num_heads: 4
    num_layers: 2
    learning_rate: 0.0003
    gamma: 0.99
    gae_lambda: 0.95
    clip_epsilon: 0.2
    value_coef: 0.5
    entropy_coef: 0.01
    max_grad_norm: 0.5
    ppo_epochs: 4
    batch_size: 15
    lr_schedule: "cosine"

  mappo:
    hidden_dim: 256
    critic_hidden_dim: 256
    learning_rate: 0.0003
    gamma: 0.99
    gae_lambda: 0.95
    clip_epsilon: 0.2
    value_coef: 0.5
    entropy_coef: 0.01
    max_grad_norm: 0.5
    ppo_epochs: 4
    batch_size: 15
    lr_schedule: "cosine"

  tcamappo:
    hidden_dim: 256
    critic_hidden_dim: 256
    history_window: 6
    critic_num_heads: 4
    critic_num_layers: 2
    critic_dropout: 0.05
    learning_rate: 0.0003
    gamma: 0.99
    gae_lambda: 0.95
    clip_epsilon: 0.2
    value_coef: 0.5
    entropy_coef: 0.01
    max_grad_norm: 0.5
    ppo_epochs: 4
    batch_size: 15
    lr_schedule: "cosine"

  dcmappo:
    hidden_dim: 256
    critic_hidden_dim: 256
    num_corridor_blocks: 2
    corridor_kernel_size: 5
    corridor_dropout: 0.05
    learning_rate: 0.0003
    gamma: 0.99
    gae_lambda: 0.95
    clip_epsilon: 0.2
    value_coef: 0.5
    entropy_coef: 0.01
    max_grad_norm: 0.5
    ppo_epochs: 4
    batch_size: 15
    lr_schedule: "cosine"

  ddpg:
    learning_rate: 0.0003
    gamma: 0.99
    buffer_size: 20000
    learning_starts: 200
    batch_size: 128
    tau: 0.005
    exploration_sigma: 0.15

  sac:
    learning_rate: 0.0003
    gamma: 0.99
    buffer_size: 20000
    learning_starts: 200
    batch_size: 128
    tau: 0.005
    ent_coef: "auto"
    target_entropy: "auto"
    target_update_interval: 1
    log_std_init: -3.0
    activation_fn: "relu"
    actor_hidden_dims: [256, 256]
    critic_hidden_dims: [256, 256]

  td3:
    learning_rate: 0.0003
    gamma: 0.99
    buffer_size: 20000
    learning_starts: 200
    batch_size: 128
    tau: 0.005
    policy_delay: 2
    exploration_sigma: 0.15
    activation_fn: "relu"
    actor_hidden_dims: [400, 300]
    critic_hidden_dims: [400, 300]

  sctd3:
    learning_rate: 0.0003
    gamma: 0.99
    buffer_size: 20000
    learning_starts: 200
    batch_size: 128
    tau: 0.005
    policy_delay: 2
    exploration_sigma: 0.15
    activation_fn: "relu"
    actor_hidden_dims: [400, 300]
    critic_hidden_dims: [400, 300]
    extractor_feature_dim: 128
    extractor_edge_hidden_dim: 16
    extractor_global_hidden_dim: 32
    extractor_spatial_blocks: 1
    extractor_kernel_size: 3