general:
  experiment_tag: 'lstm_drqn_'
  env_id: 'twcc_easy_level10_gamesize100_step50_seed9_train'
  run_test: True
  valid_env_id: 'twcc_easy_level10_gamesize10_step50_seed9_validation'
  test_env_id: ['twcc_easy_level5_gamesize10_step50_seed1_test',
                'twcc_easy_level10_gamesize10_step50_seed0_test',
                'twcc_easy_level15_gamesize10_step50_seed3_test',
                'twcc_easy_level20_gamesize10_step50_seed2_test',
                'twcc_easy_level30_gamesize10_step50_seed3_test']
#  test_env_id: ['twcc_easy_level10_gamesize10_step50_seed0_test']
  discount_gamma: 0.5
  random_seed: 42
  observation_cache_capacity: 1  # concat window of history observation, 1 means no history observations available
  experiments_dir: 'experiments/lstm_drqn'
  use_cuda: True  # disable this when running on machine without cuda
  provide_prev_action: True
  num_grids: 25
  margin: 3

  # replay memory
  history_size: 8
  update_from: 4
  replay_memory_capacity: 500000 
  replay_memory_priority_fraction: 0.25  # 0.0 to disable this
  update_per_k_game_steps: 4
  replay_batch_size: 32

  # epsilon greedy
  epsilon_anneal_epochs: 1000  # -1 if not annealing
  epsilon_anneal_from: 1.0
  epsilon_anneal_to: 0.2

  # counting reward
  revisit_counting: True
  revisit_counting_lambda_anneal_from: 1.0
  revisit_counting_lambda_anneal_epochs: -1  # -1 if not annealing
  revisit_counting_lambda_anneal_to: 0.0

training:
  scheduling:
    batch_size: 10
    test_batch_size: 10
    epoch: 3000
    model_checkpoint_path: 'saved_models/model1.pt'
    logging_frequency: 20

  optimizer:
    step_rule: 'adam'  # adam, sgd
    learning_rate: 0.001
    clip_grad_norm: 5

model:
  lstm_dqn:
    embedding_size: 20
    encoder_rnn_hidden_size: [100]
    action_scorer_hidden_dim: 64
    dropout_between_rnn_layers: 0.