rl/configs/recurrent.yaml

# Phase 3 -- Recurrent PPO (LSTM) across all tasks
# Uses existing 28-action design.

hyperparameters:
  learning_rate: 0.0001
  n_steps: 512
  batch_size: 128
  n_epochs: 10
  gamma: 0.995
  gae_lambda: 0.95
  clip_range: 0.2
  ent_coef: 0.002
  vf_coef: 0.5
  max_grad_norm: 0.5
  net_arch: [256, 256]
  lstm_hidden_size: 128
  n_lstm_layers: 1
  shared_lstm: false
  enable_critic_lstm: true
  recurrent_seq_len: 16

curriculum:
  stage1_end_frac: 0.15
  stage2_end_frac: 0.50
  stage3_weights: [0.15, 0.35, 0.50]

training:
  total_timesteps: 600000
  n_envs: 4
  seed: 42
  warm_start_from: "results/best_model/phase2_final"
  transfer_flat_weights: true
  eval_task_id: "mixed_urgency_medium"
  eval_freq: 2048
  n_eval_episodes: 3

target_scores:
  district_backlog_easy: 0.82
  mixed_urgency_medium: 0.75
  cross_department_hard: 0.68
  average: 0.82