Hello and good day.
Recently I found out about ML-Agents and that it’s possible to make AI for games with them and I thought maybe I could use them for my game’s AI but I’m not sure if they fit my needs.
What I need is some sort of a typical AI for a FPS game, a state based system with some behaviors like wandering, attacking, covering, etc for each enemy and probably a manager to prevent enemies from attacking all at once.
Is it possible to create such system using ML-Agents and does it worth the effort? or should I just write a state machine?
Thanks for you help!
It’s definitely possible with ML-Agents. You could do it in one of two ways - fully end-to-end with ML-Agents (agents just learn how to beat the enemy, but with little control as to what they’re doing) or in combination with a state machine. You could for instance train different policies for wandering, attacking, etc, and pick between them using a state machine. We’re working on a feature to enable multiple behaviors using one policy, so stay tuned for that.
Where ML-Agents is really good is for learning reactive behaviors (e.g. how to kick a physics-based ball to get it into a goal, how to aim and shoot at enemies, how to fly a plane). It’s less good at “planning” behaviors (e.g. playing chess). Hard to say whether the traditional approach or the ML approach would be easier, but it’s definitely possible.