Does game Pendulum really need take up to 1 hour to converge? about alphazero-gym HOT 2 CLOSED

I see, thanks for your answer.

from alphazero-gym.

Personally, with raw SAC, I could make pendulum converge to a range around [-500,-200)

It's the same for this repo. With a bit of hyperparameter tuning you can get to around -200 repeatedly.

1 minute on laptop of a Geforce 940M GPU

If your laptop has a GPU, I'd say it's well above average :) The one hour refers to e.g. your standard Intel built-in graphics. AlphaZero will still be a lot slower than SAC though due to the planning required at every time step. MCTS rollouts in Python are just very slow...
You can try to adapt the n_rollouts parameter in the MCTS config and see if you can lower it without affecting performance. That's the parameter affecting training throughput the most.

from alphazero-gym.