Search Unity

Question Will ML-Agents work in my RTS game?

Discussion in 'ML-Agents' started by Programmingat23505, Aug 11, 2021.

  1. Programmingat23505

    Programmingat23505

    Joined:
    Sep 15, 2020
    Posts:
    44
    I have a online multiplayer rts game I’m developing, and sometimes when you play a game the arena is empty without AI. The maps are user generated, so the ai should be able to adapt. So is ml-agents even possible in this sense?

    Thanks!
     
    Programmingat23505, Aug 11, 2021
    #1

  2. mahon123

    mahon123

    Unity Technologies

    Joined:
    Feb 9, 2021
    Posts:
    13
    A lot of ml-agents example environment are real time and require strategy to win the game. RL learns best with a well-shaped reward and self-play with previous version of the model. See 1 , 2

    Will it be able to generalize to new maps? depends, as long as the observation & action space remains the same and there's no drastic change. It's also possible to fine-tune a pre-trained model on new maps.
     
    mahon123, Aug 12, 2021
    #2

  3. Programmingat23505

    Programmingat23505

    Joined:
    Sep 15, 2020
    Posts:
    44
    I think handling new maps is a smaller hurdle, but where do you even start? I get making ml-agents drive a car or touch a button/cut and dry actions. How would you start with placing buildings and driving units? I currently use nav-mesh agents for navigation!

    Thanks!
     
    Programmingat23505, Aug 12, 2021
    #3

  4. Programmingat23505

    Programmingat23505

    Joined:
    Sep 15, 2020
    Posts:
    44
    Could one of the ml-agents spawn the buildings and the actual units are their own agents? If so then you could train the main ai to strategize on building placement and walls.

    Lastly, can I would use multi-agent training, but asymmetric because each unit and the team leader is different.
     
    Last edited: Aug 12, 2021
    Programmingat23505, Aug 12, 2021
    #4