# This can reach 18-19 reward within 10 minutes on a Tesla M60 GPU (e.g., G3 EC2 node)
# with 32 workers and 10 envs per worker. This is more efficient than the non-vectorized
# configuration which requires 128 workers to achieve the same performance.
pong-impala-vectorized:
    env: PongNoFrameskip-v4
    run: IMPALA
    config:
        sample_batch_size: 50
        train_batch_size: 500
        num_workers: 32
        num_envs_per_worker: 10