pendulum_crr:
    env: 'Pendulum-v1'
    run: CRR
    stop:
      evaluation/episode_reward_mean: -200
      training_iteration: 500
    config:
        input:
          - 'tests/data/pendulum/pendulum_replay_v1.1.0.zip'
        framework: torch
        gamma: 0.99
        train_batch_size: 1024
        critic_hidden_activation: 'relu'
        critic_hiddens: [256, 256]
        critic_lr: 0.0003
        actor_hidden_activation: 'relu'
        actor_hiddens: [256, 256]
        actor_lr: 0.0003
        actions_in_input_normalized: True
        clip_actions: True
        # Q function update setting
        twin_q: True
        target_update_grad_intervals: 1
        tau: 0.0001
        # evaluation
        evaluation_config:
            explore: False
            input: sampler
        evaluation_duration: 10
        evaluation_duration_unit: episodes
        evaluation_interval: 1
        evaluation_num_workers: 1
        evaluation_parallel_to_training: True
        # replay buffer
        replay_buffer_config:
          type: ray.rllib.utils.replay_buffers.MultiAgentReplayBuffer
          learning_starts: 0
          capacity: 100000
        # specific to CRR
        weight_type: exp
        advantage_type: max
        max_weight: 20.0
        n_action_sample: 4