pong-deterministic-rainbow:
    env: ALE/Pong-v5
    run: DQN
    stop:
        sampler_results/episode_reward_mean: 20
    config:
        # Make analogous to old v4 + NoFrameskip.
        env_config:
            frameskip: 1
            full_action_space: false
            repeat_action_probability: 0.0
        num_atoms: 51
        noisy: True
        gamma: 0.99
        lr: .0001
        hiddens: [512]
        rollout_fragment_length: 4
        train_batch_size: 32
        exploration_config:
          epsilon_timesteps: 2
          final_epsilon: 0.0
        target_network_update_freq: 500
        replay_buffer_config:
          type: MultiAgentPrioritizedReplayBuffer
          prioritized_replay_alpha: 0.5
          capacity: 50000
        num_steps_sampled_before_learning_starts: 10000
        n_step: 3
        gpu: True
        model:
          grayscale: True
          zero_mean: False
          dim: 42
        # we should set compress_observations to True because few machines
        # would be able to contain the replay buffers in memory otherwise
        compress_observations: True