config.yaml

general:
  train_data_size: 500  #-1 means unlimited games, where games are generated on the fly. In paper we use [1, 2, 10, 100, 500, -1], feel free to use any number.
  testset_path: "test_set"
  random_map: True  # If True, map size is sampled from [2, 12], if False, then map size is fixed to be 6.
  question_type: "existence"  # location, existence, attribute
  random_seed: 42  # the secret of deep learning.
  naozi_capacity: 1  # agent keeps this many history observations. In paper we do not use a size larger than 1, i.e., we do not keep history. Feel free to play with it, tho.
  use_cuda: False  # disable this when running on machine without cuda

training:
  batch_size: 1
  max_episode: 10000
  learn_start_from_this_episode: 10000
  target_net_update_frequency: 200000  # sync target net with online net per this many epochs
  max_nb_steps_per_episode: 80  # after this many steps, a game is terminated
  qa_loss_lambda: 1.0
  interaction_loss_lambda: 1.0
  optimizer:
    step_rule: 'adam'  # adam
    learning_rate: 0.00025
    clip_grad_norm: 5

evaluate:
  run_eval: True
  batch_size: 1
  max_nb_steps_per_episode: 50  # after this many steps, a game is terminated
  initial_reward: 1 # -1 for random sampling where reward ~ exp(lambda)
  silent: False # If true, prints out sufficient info and QA scores.
  eval_nb_episodes: 500 # Number of evaluation episodes [1,500]
  shuffle: False # Shuffle the episodes evaluated on

checkpoint:
  save_checkpoint: False
  save_frequency: 1000  # episode
  experiment_tag: ''
  load_pretrained: True  # during test, enable this so that the agent load your pretrained model
  load_from_tag: ''  # if you want to load from prev experiment

replay:
  discount_gamma: 0.9
  replay_memory_capacity: 500000  # adjust this depending on your RAM size
  replay_memory_priority_fraction: 0.5
  update_per_k_game_steps: 20
  replay_batch_size: 64
  multi_step: 1

distributional:
  enable: False
  atoms: 51
  v_min: -10
  v_max: 10

dueling_networks: False

double_dqn: True

epsilon_greedy:
  noisy_net: False  # if this is true, then epsilon greedy is disabled
  epsilon_anneal_episodes: 100000  # -1 if not annealing
  epsilon_anneal_from: 1.0
  epsilon_anneal_to: 0.1

valid_command_bonus_lambda: 0.  # give reward when a generated command is in TextWorld's provided admissible command list. Note we don't use this in the paper.

episodic_counting_bonus:
  revisit_counting: True
  revisit_counting_lambda_anneal_from: 0.1
  revisit_counting_lambda_anneal_episodes: -1  # -1 if not annealing
  revisit_counting_lambda_anneal_to: 0.0

model:
  use_pretrained_embedding: False
  word_embedding_size: 300
  word_embedding_trainable: True
  char_embedding_size: 32
  char_embedding_trainable: True
  embedding_dropout: 0.
  encoder_layers: 1
  aggregation_layers: 3
  encoder_conv_num: 2
  aggregation_conv_num: 2
  block_hidden_dim: 64
  n_heads: 1
  attention_dropout: 0.
  block_dropout: 0.
  action_scorer_hidden_dim: 64
  question_answerer_hidden_dim: 64