config.yaml

# goes in PufferLib top directory
train:
  seed: 1
  torch_deterministic: True
  device: cuda
  learning_rate: 0.0002 # 0.0002 # 0.000175 # 0.00015
  anneal_lr: False
  gamma: 0.998
  gae_lambda: 0.95
  num_minibatches: 1 # 4 # 16 # 32 didn't work well? kind of slow? # 16 worked well # 4 default
  norm_adv: True
  clip_coef: 0.1
  clip_vloss: True
  ent_coef: 0.01
  vf_coef: 0.5
  max_grad_norm: 0.5
  target_kl: ~

  env_pool: True  
  verbose: True 
  data_dir: experiments
  checkpoint_interval: 200
  cpu_offload: True
  pool_kernel: [0]
  bptt_horizon: 16
  vf_clip_coef: 0.1
  compile: True
  compile_mode: reduce-overhead

sweep:
  method: random
  name: sweep
  metric:
    goal: maximize
    name: episodic_return
  # Nested parameters name required by WandB API
  parameters:
    train:
      parameters:
        learning_rate: {
          'distribution': 'log_uniform_values',
          'min': 1e-4,
          'max': 1e-1,
        }
        batch_size: {
          'values': [128, 256, 512, 1024, 2048],
        }
        batch_rows: {
          'values': [16, 32, 64, 128, 256],
        }
        bptt_horizon: {
          'values': [4, 8, 16, 32],
        }

# fastest badge 1 ever: (lr: 2.0e-4, anneal=True, num_envs=150, envs_per_worker=2, envs_per_batch=60, update_epochs=3, batch_size=32768, batch_rows=64)
pokemon_red:
  package: pokemon_red
  train:
    total_timesteps: 1_000_000_000
    num_envs: 72 # 96 # 144 # 72
    envs_per_worker: 1 # 1
    envs_per_batch: 18 # 48 is crap w/ 96 env # 24
    update_epochs: 3 # 3
    gamma: 0.998
    batch_size: 49152 # 65536 # 49152
    batch_rows: 128 # 128
  env:
    name: pokemon_red
pokemon-red:
  package: pokemon_red
pokemonred:
  package: pokemon_red
pokemon:
  package: pokemon_red
pokegym:
  package: pokemon_red