First of all I have some experience in game development (Java) and with machine learning (TensorFlow in python).
I want to integrate my own game in OpenAI as an environment, to test different RL Agents.
My current game:
My game is a bit like the first few FinalFantasy games, so it’s turn-based and the user can place a character on an board, move this character and attack the AI.
The current AI is decision based, so really bad.
The game is implemented in java and uses JSON to communicate through an websocket to an browser.
I think that doesn’t matter, because I can change it, so no login is needed for openai.
But anywhere the login and registration runs on an webserver in node.
I want to create an connection between my game and openai, so that my RL agent will play against some other RL agent. So I would like to use these and not my old hard coded AI.
As soon as I found an agent which is good enough I want to use it as my new AI for this game. This AI will than play against real users.
Why agent vs. agent?
Because I would not to play versus the agent (for training may take long time) on my own or the new Agent against the current AI.
Here my questions:
I want to ask you if it is possible to use this game with its existing architecture on openai, I think universe will fit this.
Secondly is it possible to use RL Agent vs RL Agent for training, is so how can I do it.
Or if not which other possibilities out their to train an AI which should be an AI in the further game and not like in the most environments, an AI which plays for an real player.
I hope you can help me.
If something is unclear don't hesitate to ask me.
Thanks in advance