recomm-sys001-ppo:
    env: ray.rllib.examples.env.recommender_system_envs.RecommSys001
    run: PPO
    stop:
        #evaluation/episode_reward_mean: 48.0
        timesteps_total: 200000
    config:
        framework: tf

        metrics_num_episodes_for_smoothing: 1000

        # Env c'tor kwargs:
        env_config:
              # Number of different categories a doc can have and a user can
              # have a preference for.
              num_categories: 2
              # Number of docs to choose (a slate) from each timestep.
              num_docs_to_select_from: 10
              # Slate size.
              slate_size: 1
              # Re-sample docs each timesteps.
              num_docs_in_db: 100
              # Re-sample user each episode.
              num_users_in_db: 100
              # User time budget (determines lengths of episodes).
              user_time_budget: 60.0

        # Larger networks seem to help (large obs/action spaces).
        model:
          fcnet_hiddens: [256, 256]

        # Larger batch sizes seem to help (more stability, even with higher lr).
        #train_batch_size: 32

        #num_workers: 2
        #num_gpus: 0

        #lr_choice_model: 0.002
        #lr_q_model: 0.002

        #target_network_update_freq: 500
        #tau: 1.0

        # Evaluation settings.
        evaluation_interval: 1
        evaluation_num_workers: 4
        evaluation_duration: 200
        evaluation_duration_unit: episodes
        evaluation_parallel_to_training: true