cartpole_crr:
    env: 'CartPole-v0'
    run: CRR
    stop:
      evaluation/episode_reward_mean: 200
      training_iteration: 100
    config:
        input:
          - 'tests/data/cartpole/large.json'
        framework: torch
        gamma: 0.99
        train_batch_size: 2048
        critic_hidden_activation: 'tanh'
        critic_hiddens: [128, 128, 128]
        critic_lr: 0.0003
        actor_hidden_activation: 'tanh'
        actor_hiddens: [128, 128, 128]
        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: 0.0005
        # 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: bin
        advantage_type: max
        max_weight: 20.0
        n_action_sample: 4