# PPO tuned for MJX (1024+ parallel envs on GPU).
# With 1024 envs, each timestep collects 1024 samples, so total_timesteps
# can be much lower than the CPU config.

hidden_sizes: [128, 128]
total_timesteps: 300000       # 300K × 1024 envs ≈ 307M env steps
rollout_steps: 1024           # PPO batch = 1024 envs × 1024 steps = 1M samples
learning_epochs: 4
mini_batches: 32              # keep mini-batch size similar to CPU config (~32K)
discount_factor: 0.99
gae_lambda: 0.95
learning_rate: 0.001          # ~3x higher LR for 16x larger batch (sqrt scaling)
clip_ratio: 0.2
value_loss_scale: 0.5
entropy_loss_scale: 0.05
log_interval: 100             # log more often (shorter run)
checkpoint_interval: 10000

record_video_every: 10000

# ClearML remote execution (GPU worker)
remote: false