# This file runs RLlib algorithm learning tests for select algorithms on torch.
# It is suggested to run these on a single g3.16xlarge or p3.16xl node
# in a DLAMI / tensorflow_p36 env.
# Note: RL runs are inherently high variance, so you'll have to check to
# see if the rewards reached seem reasonably in line with previous results.
# You can find the reference results here:
# https://github.com/ray-project/ray/tree/master/release/release_logs

a2c-torch-atari:
    env: BreakoutNoFrameskip-v4
    run: A2C
    num_samples: 2
    stop:
        time_total_s: 3600
    config:
        framework: torch
        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-dqn-torch-atari:
    env: BreakoutNoFrameskip-v4
    run: APEX
    num_samples: 2
    stop:
        time_total_s: 3600
    config:
        framework: torch
        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-torch-atari:
    env: BreakoutNoFrameskip-v4
    run: DQN
    num_samples: 2
    stop:
        time_total_s: 3600
    config:
        framework: torch
        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-torch-atari:
    env: BreakoutNoFrameskip-v4
    run: IMPALA
    num_samples: 2
    stop:
        time_total_s: 3600
    config:
        framework: torch
        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-torch-atari:
    env: BreakoutNoFrameskip-v4
    run: PPO
    num_samples: 2
    stop:
        time_total_s: 3600
    config:
        framework: torch
        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-torch-halfcheetah-pybullet:
    env: HalfCheetahBulletEnv-v0
    run: SAC
    num_samples: 2
    stop:
        time_total_s: 3600
    config:
        framework: torch
        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