I’m thinking about hardcoding “engrave Elbereth with finger” as a single action available to my agent.
Two questions for the organizers:
Is this allowed? I believe its mentioned in the original NLE paper as a possibility but you would expect authors to specify that they did it (and other hardcoding) - how does that apply to this competition? Would hardcoding it as an action available to an otherwise standard deep RL model move such a model into the “not using a neural network” track?
If it is allowed, what would be the preferred/allowed level to implement it at? Currently I’ve extended the base NLE gym.Env (in nle/env/base.py) to check for a specific action id (I manually expanded the discrete action space by 1) in step and then run the hardcoded steps to engrave - but this feels pretty hacky. I’m guessing the testing environment won’t allow modifications to the nle package itself so any changes will have to be in the challenge env wrappers.py. Is that correct?
Thanks, and thanks also for the fun competition!