interest-evolution-recsim-env-slateq:
    env: ray.rllib.examples.env.recsim_recommender_system_envs.InterestEvolutionRecSimEnv
    run: SlateQ
    stop:
        episode_reward_mean: 200.0
        timesteps_total: 60000
    config:
        # SlateQ only supported for torch so far.
        framework: torch

        # RLlib/RecSim wrapper specific settings:
        env_config:
            config:
                # Each step, sample 100 documents using the env-internal
                # document sampler model (a logic that creates n documents to select
                # the slate from).
                resample_documents: false
                num_candidates: 10
                # How many documents to recommend (out of `num_candidates`) each
                # timestep?
                slate_size: 2
                # Should the action space be purely Discrete? Useful for algos that
                # don't support MultiDiscrete (e.g. DQN).
                # SlateQ handles MultiDiscrete action spaces.
                convert_to_discrete_action_space: false
                seed: 42

        hiddens: [256, 256]

        num_workers: 0
        num_gpus: 0

        lr_choice_model: 0.003
        lr_q_model: 0.003
        rollout_fragment_length: 4
        exploration_config:
            epsilon_timesteps: 50000

        target_network_update_freq: 800