# Can expect improvement to -140 reward in ~300-500k timesteps.
pendulum-ppo:
    env: ray.rllib.examples.env.transformed_action_space_env.TransformedActionPendulum
    run: PPO
    stop:
        episode_reward_mean: -500
        timesteps_total: 400000
    config:
        # Works for both torch and tf.
        framework: tf

        # Test, whether PPO is able to learn in "distorted" action spaces.
        env_config:
            config:
                low: 300.0
                high: 500.0

        normalize_actions: true
        clip_actions: false
        vf_clip_param: 10.0
        num_envs_per_worker: 20
        lambda: 0.1
        gamma: 0.95
        lr: 0.0003
        train_batch_size: 512
        sgd_minibatch_size: 64
        num_sgd_iter: 6
        observation_filter: MeanStdFilter