# This configuration can expect to reach 2000 reward in 150k-200k timesteps
halfcheetah-ddpg:
    env: HalfCheetah-v2
    run: DDPG
    stop:
        episode_reward_mean: 2000
        time_total_s: 5400 # 90 minutes
    config:
        # === Model ===
        actor_hiddens: [64, 64]
        critic_hiddens: [64, 64]
        n_step: 1
        model: {}
        gamma: 0.99
        env_config: {}

        # === Exploration ===
        exploration_should_anneal: True
        schedule_max_timesteps: 100000
        timesteps_per_iteration: 1000
        exploration_fraction: 0.1
        exploration_final_scale: 0.02
        exploration_ou_noise_scale: 0.1
        exploration_ou_theta: 0.15
        exploration_ou_sigma: 0.2
        target_network_update_freq: 0
        tau: 0.001

        # === Replay buffer ===
        buffer_size: 10000
        prioritized_replay: True
        prioritized_replay_alpha: 0.6
        prioritized_replay_beta: 0.4
        prioritized_replay_eps: 0.000001
        clip_rewards: False

        # === Optimization ===
        actor_lr: 0.001
        critic_lr: 0.001
        use_huber: False
        huber_threshold: 1.0
        l2_reg: 0.000001
        learning_starts: 500
        sample_batch_size: 1
        train_batch_size: 64

        # === Parallelism ===
        num_workers: 0
        num_gpus_per_worker: 0
        per_worker_exploration: False
        worker_side_prioritization: False

        # === Evaluation ===
        evaluation_interval: 5
        evaluation_num_episodes: 10