Make enemies smarter with time using machine learning

Questions & Answers
1

Hi, I’m considering using ml-agents to make an enemy ai that should chase the player and become adaptive to player movements/actions over time (e.g. dodge player bullets, corner the player) and I’m not sure how to begin making this.

Thanks for the help!,

2

Hey there and welcome to the forum,

assuming that this is your first real project with ML-Agents: Please don’t start with this. it will only bring you frustration.

Why? Because in the end this idea can (imo) only really bear fruits in a way where you might see that the AI has learnt something when the problem (so here “playing the game”) is sufficiently complex. But then again solving a complex problem is difficult for AI. Even more for Reinforcement Learning as you need a TON of learning time.

This is also something that you might not be aware of: Learning times are long (can be hours) even for simple to medium problems. So when you have the idea that the AI learns to adjust to player behaviour in a matter of just a few games then you will have to find a way to basically pre-learn different possible behaviours that a player could have and then try to swithc these on and off

Now at this point you might realize that this solution sounds more like a normal state-machine - and it probably really is.

In the end there is a reason why basically noone is doing this right? (only exception that comes to mind is “hello neighbour”)

So to sum it up: If this is your first ML-Agents project - don’t start here. Do some small fun project where the AI has to learn to play flappy birds or something.

This will give you the experience you need to tackle a more complex issue (let’s say have an AI solve a maze, then have it solve any maze based on the model you created - this is more difficult than it sounds)

After that you will have more knowledge to decide if what you want to do can be done or not.
This is basically all i can help you with given the information that is given. For more specific help (for example how to design observation and action space you’d have to provide way way more details on what this game would look like.