How to set seed in your project?

[zhuzil]

Congratulations on the outstanding paper :)

I would like to know how to set the seed in your project. I only saw the setting of the environment seed in AtariEnv and DeepMindControlEnv, but I couldn't find the calls for the global seed such as torch.manual_seed() or torch.cuda.manual_seed()?

Is it sufficient to only set the environment seed?

[burchim]

Hi, thanks for your interest!

We did not set the seed for experiments, simply relying on random seeds by default for PyTorch, the environments and all other libraries.

But if you wish to have deterministic runs with controllable seeds, setting the PyTorch seed will also be necessary for the sampling randomness of the replay buffer.

Concerning AtariEnv, you will also need to set the numpy seed with numpy.random.seed(seed) for the inner number generator.

Concerning Deep Mind Control, the correct way to set environment seed is actually to pass the task_kwargs={'random': seed} parameter to suite.load(). The Deep Mind Control set_seed() function needs to be updated !