# Given a SAC-generated offline file generated via:
# rllib train -f tuned_examples/sac/pendulum-sac.yaml --no-ray-ui

# Pendulum CQL can attain ~ -300 reward in 10k from that file.
pendulum-cql:
    env: Pendulum-v0
    run: CQL
    stop:
        episode_reward_mean: -300
        #timesteps_total: 10000
    config:
        # Works for both torch and tf.
        framework: tf
        # Use one or more offline files or "input: sampler" for online learning.
        input: ["/your/json/file/here"]
        horizon: 200
        soft_horizon: true
        Q_model:
          fcnet_activation: relu
          fcnet_hiddens: [256, 256]
        policy_model:
          fcnet_activation: relu
          fcnet_hiddens: [256, 256]
        tau: 0.005
        target_entropy: auto
        no_done_at_end: true
        n_step: 3
        rollout_fragment_length: 1
        prioritized_replay: false
        train_batch_size: 256
        target_network_update_freq: 1
        timesteps_per_iteration: 1000
        learning_starts: 256
        optimization:
          actor_learning_rate: 0.0003
          critic_learning_rate: 0.0003
          entropy_learning_rate: 0.0003
        num_workers: 0
        num_gpus: 1
        clip_actions: False
        normalize_actions: true
        evaluation_num_workers: 1
        evaluation_interval: 1
        metrics_smoothing_episodes: 5
        bc_iters: 0
        # Evaluate in an actual environment.
        evaluation_config:
            input: sampler
            explore: False