liars_sp.yaml

defaults:
  - optimizer: adam
seed: 0
decrease_lr_every: 400
decrease_lr_times: 2
grad_clip: 5.0
env:
  num_dice: 1
  num_faces: 4
  random_action_prob: 0.25
  sample_leaf: true
  subgame_params:
    num_iters: 1024
    max_depth: 2
    linear_update: true
exploit: true
selfplay:
  network_sync_epochs: 1
  dump_dataset_every_epochs: 200
  threads_per_gpu: -1
  models_per_gpu: 1
  cpu_gen_threads: 60
train_gen_ratio: 4
task: selfplay
loss: huber
min_buffer_to_send: 2500
max_epochs: 10000
model:
  name: Net2
  kwargs:
    n_hidden: 256
    n_layers: 2
    use_layer_norm: true
create_validation_set_every: 100
data:
  train_epoch_size: 25600
  train_batch_size: 512
replay:
  capacity: 2000000
  alpha: 1.0
  beta: 1.0
  prefetch: 8
  use_priority: false