Can you use Sentis to create dynamic AI mobs?

For example, I assume everyone is familiar with League of Legends. So would it be possible to use Sentis to create a AI opponent, powered by a ML model, which plays against players and trains/improves while playing the game?

I think that the ML-Agents package is better suited for this purpose, but you’ll probably end up with an AI that’s not much fun to play against. Creating a balanced AI will be difficult.

Well, from what I understood from the ML-Agents package, you have to train them offline and then activate it during an actual game play session. I want the AI to discover continuously while you play and I disagree that it would make it not fun to play against because I would very much want to play against a dynamic AI which changes it behavior as you play, rather than one that just follow static algorithms.

While Unity Sentis has the potential for creating dynamic AI mobs, it’s currently more suited for experienced developers with access to pre-trained AI models and the ability to optimize for performance. For simpler dynamic enemy behavior, alternatives like behavior trees might be more practical.

1 Like

Hi Shakor. What you are referring to is known as “train-on-device” using reinforcement learning. While we don’t currently have examples of this, we are looking into it. (We currently don’t have a time schedule for this, though). I believe a very basic algorithm for this was used in the 2001 game Black and White which coincidentally was programmed by the current head of Google Deep Mind.

An alternative method for what you describe is to train a model to be a very good player, but then adjust the difficulty by increasing the randomness of the opponent. This is how we achieved a difficulty setting with our Othello sample.