slateq-interest-evolution-recsim-env:
    env: ray.rllib.examples.env.recommender_system_envs_with_recsim.InterestEvolutionRecSimEnv
    run: SlateQ
    pass_criteria:
        episode_reward_mean: 160.0
        timesteps_total: 300000
    stop:
        time_total_s: 7200
    config:
        # increase num sampling workers for faster sampling.
        num_workers: 12
        # RLlib/RecSim wrapper specific settings:
        env_config:
            # Env class specified above takes one `config` arg in its c'tor:
            config:
                # Each step, sample `num_candidates` documents using the env-internal
                # document sampler model (a logic that creates n documents to select
                # the slate from).
                resample_documents: true
                num_candidates: 50
                # 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 or Bandits).
                # SlateQ handles MultiDiscrete action spaces.
                convert_to_discrete_action_space: false
                seed: 0

        num_gpus: 1

        exploration_config:
            warmup_timesteps: 20000
            epsilon_timesteps: 70000

        replay_buffer_config:
            capacity: 500000

        lr: 0.00025

        metrics_num_episodes_for_smoothing: 200