Using Action-based to create repeatable reward only when the action is executed

microsoft / TextWorld

TextWorld is a sandbox learning environment for the training and evaluation of reinforcement learning (RL) agents on text-based games.

Other

1.23k stars 188 forks source link

Using Action-based to create repeatable reward only when the action is executed #299

Open lanhoang712 opened 2 years ago

lanhoang712 commented 2 years ago

Hi @MarcCote we have a follow-up question on issue #296 . Is there a way to make the reward repeatable but only occur once with the action? The problem when I tested the "eat apple", "take apple" approach is that the reward keeps occurring as long as the apple is in the inventory (i.e. until "drop apple"). We want the quest to be repeatable so that the agent is rewarded with each time it takes the apple, but only with the action "take". Is there a way to do that?

MarcCote commented 2 years ago

That's a good point. You are right that's an issue with the current system. There's a stale pull request implementing what you need (see #235). I'll try to spend some time next to finish it. I remember it being pretty close to completion.