Items in RL Config File

• cluster: dictionary of 'job name': 'addr:port' pairs

• datapool:

  • pool_size: max number of moves in the pool
  • start_data_size: min number of moves needed to start training
  • conn_num: max number of concurrent connections with self play games
  • load_prev: whether to load previous self play data
  • store_path: path to store the self play data. unset to disable

• evaluator:

  • num_games: total number of games for calculating the winning rate
  • num_worker: max number of concurrent games
  • gameplay:
    • dirichlet_before: the number of actions using dirichlet noise at the beginning of each game
    • log_iter: interval of printing log of one game
    • max_turn: max number of turns of one game
    • player:
      • max_playout: max number of playout of MCTS

• selfplay:

  • num_worker: max number of concurrent games
  • remote_port: port to receive selfplay data from remote computers. used by master
  • remote_update_port: port to receive requests of updating model. used by slaves
  • gameplay:
    • dirichlet_before: the number of actions using dirichlet noise at the beginning of each game
    • log_iter: interval of printing log of one game
    • max_turn: max number of turns of one game
    • player:
      • max_playout: max number of playout of MCTS

• optimizer:

  • num_ckpt: interval of saving model and sending model to evaluator
  • num_steps: total number of training steps
  • batch_size: batch size
  • num_gpu: number of gpus. must set properly
  • num_log: interval of printing log of training
  • num_eval: interval of evaluating the model on validation set
  • job: job name (in the cluster spec)
  • load_path: path of pretrained model. unset to disable
  • log_dir: path to store Tensorboard event file
  • eval_data_path: path of data for validation. unset to disable
  • train_val_test: partition of data. a list of [train_percentage, val_percentage, test_percentage]. only val set will be used
  • eval_batch_size: batch size of validation

• chal:

  • max_batch_size: max batch size of a single run
  • num_gpu: number of gpus. must set properly
  • job: job name (in cluster spec)

• best:

  • max_batch_size: max batch size of a single run
  • num_gpu: number of gpus. must set properly
  • job: job name (in cluster spec)
  • load_path: path of pretrained model. unset to disable