parametric-item-reco-env-slateq:
    env: ray.rllib.examples.env.bandit_envs_recommender_system.ParametricItemRecoEnv
    run: SlateQ
    stop:
        #evaluation/sampler_results/episode_reward_mean: 48.0
        timesteps_total: 200000
    config:
        # SlateQ only supported for torch so far.
        framework: torch

        metrics_num_episodes_for_smoothing: 200

        exploration_config:
            temperature: 0.7

        # Env c'tor kwargs:
        env_config:
            config:
                slate_q: true
                num_users: 50
                num_items: 1000
                num_candidates: 50
                slate_size: 1
                feature_dim: 16

        grad_clip: 10.0
        #double_q: false
        #slateq_strategy: MYOP

        # Larger networks seem to help (large obs/action spaces).
        hiddens: [512, 512]

        # Larger batch sizes seem to help (more stability, even with higher lr).
        train_batch_size: 64

        num_workers: 0
        num_gpus: 0

        lr_choice_model: 0.01
        lr_q_model: 0.01

        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