# PPO defaults — sized for the CPU MuJoCo runner (64 parallel envs).
# 128 rollout steps × 64 envs ≈ 8K samples per update.

hidden_sizes: [256, 256]
total_timesteps: 500000          # × 64 envs = 32M env steps
rollout_steps: 128
learning_epochs: 5
mini_batches: 4
discount_factor: 0.99
gae_lambda: 0.95
learning_rate: 0.0003
clip_ratio: 0.2
value_loss_scale: 0.5
entropy_loss_scale: 0.01
kl_threshold: 0.01               # KL-adaptive LR; 0 = fixed learning rate
log_interval: 1000
checkpoint_interval: 50000

initial_log_std: -0.5
min_log_std: -4.0
max_log_std: 2.0

record_video_every: 10000

# History encoder output dim — the window size itself comes from
# runner.history_length (single source of truth).
embedding_dim: 32

# ClearML remote execution (GPU worker)
remote: false

# ── HPO search ranges ────────────────────────────────────────────────
# Read by scripts/hpo.py — ignored by TrainerConfig during training.
hpo:
  learning_rate: {min: 0.00005, max: 0.001}
  clip_ratio: {min: 0.1, max: 0.3}
  discount_factor: {min: 0.98, max: 0.999}
  gae_lambda: {min: 0.9, max: 0.99}
  entropy_loss_scale: {min: 0.0001, max: 0.1}
  value_loss_scale: {min: 0.1, max: 1.0}
  learning_epochs: {min: 2, max: 8, type: int}
  mini_batches: {values: [2, 4, 8, 16]}