constraint on action space somehow that each action can be chosen only once

Hi everyone. I am trying to solve a puzzle game with reinforcement learning and want to know if there is a way to constraint the action space or not. The problem is that I have 19 empty space in this game which has to be filled with specific numbers. Action for the agent is to chose these 19 empty spaces. Therefore if one space is chosen in the last step, it can not be chosen again, e.g. it chose space number 2 now, so it only has 18 empty spaces and consicuently 18 actions left because a filled space can not be filled again. Is there a way to do this without reward function?

Answer 1

[Emmanouil Tzorakoleftherakis]

Unfortunately, this is not possible as of R2022b out of the box. You would need to create a custom agent to implement this behavior.

https://www.mathworks.com/help/reinforcement-learning/ug/custom-agents.html