MK8 reinforcement learning

[dlibzh]

Hey! I started learning reinforcement learning and neural networks about 5-6 days ago, just for fun and out of curiosity because AI fascinates me. I'm trying to create an AI capable of completing a race in Mario Kart 8.

I initially tested image recognition, but it seems quite complex and messy for my amateur level. I'm mainly relying on forums, videos, and ChatGPT for help. Now, I’d like to shift towards reading real-time game data instead.

I’m thinking of using an MQTT broker, but my main issue is that I can't find the addresses for these values in Cheat Engine.
I tried scanning the game's values with Cheat Engine and comparing the changes as I moved through the race, but it's pretty random haha, I’m struggling quite a bit. I found a similar project with a Gym environment on GitHub, but it only works with MK8 version 1.7.1. Unfortunately, I haven’t been able to find that update on emulator forums or elsewhere.

So, I’m asking for help here—has anyone here already found these kinds of addresses or values for version 3.0.3? Specifically, I'm looking for movement speed, completed laps, and XYZ position of the player. If anyone has any tips on how to find them, that would be greatly appreciated!

thanks for your time

 

[Prof0nWear]

I remember a good video by StackSmashing on YouTube for this.

Enjoy.

 

[dlibzh]

I remember a good video by StackSmashing on YouTube for this.

Enjoy.

Yup, thanks! I had never come across this video before, even though I found dozens of them, but I hadn’t heard about TensorKart. Well, I started modifying everything to make it work on an emulator and MK8. The problem was that training would have taken hours since TensorFlow was using my CPU. And since TensorFlow, starting from version 2.10, no longer uses the GPU on Windows, I had to spend 6 hours figuring out how to get the GPU to work with TF.

I ended up creating a WSL2 (Ubuntu) environment to finally make it work, installing the right versions of CUDA, cuDNN, and Python. I got a bit sidetracked, but seeing that it’s finally working now is pretty satisfying. Training will be faster and more efficient—I just hope it was worth the effort!

Now I still have a lot of work to do to set everything up, and I hope to see some results. I'll come back here to post them if anyone's interested!