How to make AI?

If you would like to incorporate strategy and planning into the AI, there are many methods. All of them get more difficult if the AI have to work together.

I just finished CS-6600 (Intelligent systems) in which I made a sword fighting AI here on Roblox. The method, Q-Learning, is applicable for your AIs as well. If you have maps that make some positions more favorable then you may want to augment the Q-Learner with a neural network to help predict state values. The reasons why are explained in a report I put together. Here is a video of the AI training:

Here is the repo with all the source code:

Here is the class report (put together in 30 minutes, don’t tell my professor!) about what issues I ran into and what I learned:
https://drive.google.com/file/d/1MbCnC22bmX5uQZtBNHx1BKECHM_Sygu0/view?usp=sharing
And here is an article I found useful to learn about Q-learning:

Lastly, here is the place file:
AI.rbxl (80.5 KB)

To summarize, Q-Learning learns the best “policy”, for a given state (at a specific location on a map, enemy’s position, reload times, fuel, health, ext) what is the best action to take? It learns the best actions by trial and error over time. This AI was trained with a fairly small number of states so learns really fast. Other AIs may require more training or even a neural network to estimate state values. The long-term planning comes in via the propagation of rewards back along the states preceding it.

26 Likes