rl/configs/curriculum_tuned.yaml

# Phase 2 (tuned) -- curriculum continuation from existing Phase 2 checkpoint
# Minimal tuning pass: no architecture changes, 28-action design unchanged.

hyperparameters:
  learning_rate: 0.0001
  n_steps: 1024
  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]

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

training:
  total_timesteps: 300000
  n_envs: 4
  seed: 42
  warm_start_from: "results/best_model/phase2_final"
  eval_task_id: "mixed_urgency_medium"
  eval_freq: 2048
  n_eval_episodes: 3

target_scores:
  district_backlog_easy: 0.82
  mixed_urgency_medium: 0.72
  cross_department_hard: 0.60
  average: 0.75