rl/configs/ppo_easy_aggressive.yaml

# Phase 1 -- Aggressive PPO tuning (benchmark unchanged)
# Use when baseline Phase 1 plateaus around ~0.55-0.58 grader score.
#
# Example:
#   python -m rl.train_ppo --phase 1 --task district_backlog_easy --timesteps 300000 --n_envs 4 --seed 42 --phase1-config rl/configs/ppo_easy_aggressive.yaml
#
# Notes:
# - Keeps env/grader/task unchanged.
# - Focuses on longer-horizon credit assignment + lower exploration noise.

hyperparameters:
  learning_rate: 0.0001
  n_steps: 1024
  batch_size: 256
  n_epochs: 15
  gamma: 0.995
  gae_lambda: 0.98
  clip_range: 0.15
  ent_coef: 0.001
  vf_coef: 0.7
  max_grad_norm: 0.5
  net_arch: [256, 256, 128]

training:
  total_timesteps: 300000
  n_envs: 4
  seed: 42
  eval_freq: 16384
  n_eval_episodes: 3
  grader_eval_freq_multiplier: 2
  enable_eval_callback: true
  progress_bar: false
  model_verbose: 0
  callback_verbose: 0

target_scores:
  district_backlog_easy: 0.65