jax_mcts.yml

wandb:
  use: true
  work_dir: runs
  layerwise_logging: true
  log_interval: 200
  validation_interval: 1250
use_cuda: true
scale_observation: false
reposition: false
gumbel_scale: 0.3
version_string: 'mcts'
net_type: custom_dense

dimension: 3
max_num_points: 5
max_length_game: 40
max_value: 20
max_grad_norm: 0.5

eval_batch_size: 512
num_evaluations: 100
num_evaluations_as_opponent: 100
eval_on_cpu: false
max_num_considered_actions: 10
#rollout process repeats `rollout_size//eval_batch_size` amount of times
discount: 0.98

host:
  batch_size: 128
  optim:
    name: 'adam'
    args: # Pass optimizer parameters here
      learning_rate: 0.001  #
    # Learning rate scheduling is built-in (by schedule function and `chain`), so we do not specify them.
  #net_arch: [256, 256, 256, 256, 256, 256, 256, 256]
  net_arch: [128, 128, 128, 128, 128, 128, 128, 128]


agent:
  batch_size: 128
  optim:
    name: 'adam'
    args: # Pass optimizer parameters here
      learning_rate: 0.001
    # Learning rate scheduling is built-in (by schedule function and `chain`), so we do not specify them.
  #net_arch: [256, 256, 256, 256, 256, 256, 256, 256]
  net_arch: [128, 128, 128, 128, 128, 128, 128, 128]