ray/python/ray/rllib/tuned_examples/mountaincarcontinuous-ddpg.yaml

# This configuration can expect to reach 90 reward in 10k-20k timesteps
mountaincarcontinuous-ddpg:
    env: MountainCarContinuous-v0
    run: DDPG
    stop:
        episode_reward_mean: 90
        time_total_s: 600 # 10 minutes
    config:
        # === Model ===
        actor_hiddens: [32, 64]
        critic_hiddens: [64, 64]
        n_step: 3
        model: {}
        gamma: 0.99
        env_config: {}

        # === Exploration ===
        schedule_max_timesteps: 100000
        timesteps_per_iteration: 1000
        exploration_fraction: 0.4
        exploration_final_eps: 0.02
        noise_scale: 0.75
        exploration_theta: 0.15
        exploration_sigma: 0.2
        target_network_update_freq: 0
        tau: 0.01

        # === Replay buffer ===
        buffer_size: 50000
        prioritized_replay: False
        prioritized_replay_alpha: 0.6
        prioritized_replay_beta: 0.4
        prioritized_replay_eps: 0.000001
        clip_rewards: False

        # === Optimization ===
        actor_lr: 0.0001
        critic_lr: 0.001
        use_huber: False
        huber_threshold: 1.0
        l2_reg: 0.00001
        learning_starts: 1000
        sample_batch_size: 1
        train_batch_size: 64

        # === Parallelism ===
        num_workers: 0
        num_gpus_per_worker: 0
        optimizer_class: "SyncReplayOptimizer"
        per_worker_exploration: False
        worker_side_prioritization: False