mcts_config.yml

use_tensorboard: true
log_time: true
layerwise_logging: true
use_cuda: false
scale_observation: true
version_string: 'mcts_v0'

dimension: 3
max_num_points: 10
max_value: 20  # Not sure if this impacts the game at all
max_grad_norm: 10

c_puct: 0.5
lr: 0.0000001
max_depth: 20
MSE_coefficient: 1.0

# MCTS only trains one player at a time
# If host is specified, agent will need to be passed during initialization, and vice-versa.
host:
  # Note: Simulation is always run step-by-step, but the number of batches is used when training the policy net.
  batch_size: 10
  iterations: 200,
  c_puct: 0.5,
  max_depth: 20,
  MSE_coefficient: 1.0,
  c_puct: 0.5
  lr: 0.00001
  max_depth: 20
  MSE_coefficient: 1.0

  batch_size: 1
  initial_rollout_size: 100
  max_rollout_step: 10
  optim:
    name: 'adam'
    args: # Pass optimizer parameters here
      lr: 0.000001
    lr_schedule: # (OPTIONAL) Use a scheduler on the learning rate
      mode: 'exponential'
      initial_lr: 0.001
      rate: 0.996
  er: 0.2  # Exploration rate
  er_schedule: # (OPTIONAL) Use a scheduler on the exploration rate
    mode: 'exponential'
    initial_er: 0.5
    rate: 0.996
  net_arch: [ 32, { repeat: 2, net_arch: [ 32, 'b' ] }, 16 ]

agent:
  batch_size: 256
  initial_rollout_size: 100
  max_rollout_step: 10
  c_puct: 0.5
  lr: 0.00001
  max_depth: 20
  MSE_coefficient: 1.0

  optim:
    name: 'adam'
    args:
      lr: 0.00000001
    lr_schedule:
      mode: 'exponential'
      initial_lr: 0.001
      rate: 0.996
  er: 0.2
  er_schedule:
    mode: 'exponential'
    initial_er: 0.5
    rate: 0.996
  net_arch: [ 128, 'b', { repeat: 20, net_arch: [ 256, 'b' ] }, 128, 'b' ]

replay_buffer:
  type: 'base'
  buffer_size: 0
  use_cuda: false