a2c-breakoutnoframeskip-v4:
    env: BreakoutNoFrameskip-v4
    run: A2C
    # Minimum reward and total ts (in given time_total_s) to pass this test.
    pass_criteria:
        episode_reward_mean: 50.0
        timesteps_total: 5000000
    stop:
        time_total_s: 7200
    config:
        rollout_fragment_length: 20
        clip_rewards: True
        num_workers: 5
        num_envs_per_worker: 5
        num_gpus: 1
        lr_schedule: [
            [0, 0.0007],
            [20000000, 0.000000000001],
        ]

apex-breakoutnoframeskip-v4:
    env: BreakoutNoFrameskip-v4
    run: APEX
    # Minimum reward and total ts (in given time_total_s) to pass this test.
    pass_criteria:
        episode_reward_mean: 20.0
        timesteps_total: 10000000
    stop:
        time_total_s: 7200
    config:
        double_q: false
        dueling: false
        num_atoms: 1
        noisy: false
        n_step: 3
        lr: .0001
        adam_epsilon: .00015
        hiddens: [512]
        buffer_size: 1000000
        exploration_config:
          epsilon_timesteps: 200000
          final_epsilon: 0.01
        prioritized_replay_alpha: 0.5
        final_prioritized_replay_beta: 1.0
        prioritized_replay_beta_annealing_timesteps: 2000000
        num_gpus: 1
        num_workers: 8
        num_envs_per_worker: 8
        rollout_fragment_length: 20
        train_batch_size: 512
        target_network_update_freq: 50000
        timesteps_per_iteration: 25000

dqn-breakoutnoframeskip-v4:
    env: BreakoutNoFrameskip-v4
    run: DQN
    # Minimum reward and total ts (in given time_total_s) to pass this test.
    pass_criteria:
        episode_reward_mean: 30.0
        timesteps_total: 450000
    stop:
        time_total_s: 7200
    config:
        double_q: false
        dueling: false
        num_atoms: 1
        noisy: false
        prioritized_replay: false
        n_step: 1
        target_network_update_freq: 8000
        lr: .0000625
        adam_epsilon: .00015
        hiddens: [512]
        learning_starts: 20000
        buffer_size: 1000000
        rollout_fragment_length: 4
        train_batch_size: 32
        exploration_config:
          epsilon_timesteps: 200000
          final_epsilon: 0.01
        prioritized_replay_alpha: 0.5
        final_prioritized_replay_beta: 1.0
        prioritized_replay_beta_annealing_timesteps: 2000000
        num_gpus: 0.5
        timesteps_per_iteration: 10000

impala-breakoutnoframeskip-v4:
    env: BreakoutNoFrameskip-v4
    run: IMPALA
    # Minimum reward and total ts (in given time_total_s) to pass this test.
    pass_criteria:
        episode_reward_mean: 300.0
        timesteps_total: 6000000
    stop:
        time_total_s: 3600
    config:
        rollout_fragment_length: 50
        train_batch_size: 500
        num_workers: 10
        num_envs_per_worker: 5
        clip_rewards: True
        lr_schedule: [
            [0, 0.0005],
            [20000000, 0.000000000001],
        ]
        num_gpus: 1

ppo-breakoutnoframeskip-v4:
    env: BreakoutNoFrameskip-v4
    run: PPO
    # Minimum reward and total ts (in given time_total_s) to pass this test.
    pass_criteria:
        episode_reward_mean: 50.0
        timesteps_total: 10000000
    stop:
        time_total_s: 7200
    config:
        lambda: 0.95
        kl_coeff: 0.5
        clip_rewards: True
        clip_param: 0.1
        vf_clip_param: 10.0
        entropy_coeff: 0.01
        train_batch_size: 5000
        rollout_fragment_length: 100
        sgd_minibatch_size: 500
        num_sgd_iter: 10
        num_workers: 10
        num_envs_per_worker: 5
        batch_mode: truncate_episodes
        observation_filter: NoFilter
        model:
            vf_share_layers: true
        num_gpus: 1

# Expect roughly 1000 reward after 1h on 1GPU
sac-halfcheetahbulletenv-v0:
    env: HalfCheetahBulletEnv-v0
    run: SAC
    # Minimum reward and total ts (in given time_total_s) to pass this test.
    pass_criteria:
        episode_reward_mean: 400.0
        timesteps_total: 80000
    stop:
        time_total_s: 7200
    config:
        horizon: 1000
        soft_horizon: false
        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: 1
        rollout_fragment_length: 1
        prioritized_replay: true
        train_batch_size: 256
        target_network_update_freq: 1
        timesteps_per_iteration: 1000
        learning_starts: 10000
        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_interval: 1
        metrics_smoothing_episodes: 5

# Expect roughly 1000 reward after 1h on 1GPU
# TODO: (sven) this seems to be somewhat broken on tf AND torch (?)
#  try to find older version that still works.
ddpg-halfcheetahbulletenv-v0:
    env: HalfCheetahBulletEnv-v0
    run: DDPG
    # Minimum reward and total ts (in given time_total_s) to pass this test.
    pass_criteria:
        episode_reward_mean: -100.0
        timesteps_total: 400000
    stop:
        time_total_s: 7200
    config:
        actor_hiddens: [64, 64]
        critic_hiddens: [64, 64]
        n_step: 1
        model: {}
        gamma: 0.99
        env_config: {}
        exploration_config:
            initial_scale: 1.0
            final_scale: 0.02
            scale_timesteps: 10000
            ou_base_scale: 0.1
            ou_theta: 0.15
            ou_sigma: 0.2
        timesteps_per_iteration: 1000
        target_network_update_freq: 0
        tau: 0.001
        buffer_size: 10000
        prioritized_replay: True
        prioritized_replay_alpha: 0.6
        prioritized_replay_beta: 0.4
        prioritized_replay_eps: 0.000001
        clip_rewards: False
        actor_lr: 0.001
        critic_lr: 0.001
        use_huber: False
        huber_threshold: 1.0
        l2_reg: 0.000001
        learning_starts: 500
        rollout_fragment_length: 1
        train_batch_size: 64
        num_workers: 0
        num_gpus: 1
        num_gpus_per_worker: 0
        worker_side_prioritization: False