This topic will be for everything related to the machines learning and playing video games and using Artificial Intelligence, and regarding the racing/karting category; naturally of our Mario Kart games

Check & read the source: TensorKart: self-driving MarioKart with TensorFlow





Article: Artificial Intelligence Will Soon Be Better at Mario Kart Than You Are (to see this one may be required to disable your ad blocker)

Better than us at Mario Kart??? Well, better to say good luck computer and AI before beating some SMK and MK64 World Records for example but why not ^^ and looking forward to see this evolution of playing


Also TensorKart reminds me when Super Mario Kart cross the MarI/O program = Super MarI/O Kart:



It's more than TAS so the question is: in the future do you think machines will replace humans for doing our P/WRs?

ALAKTORN wrote on 01/06/17 at 09:36:45:


Yup but I was referring to the people cheating, imagine if they use this AI instead of really playing and claim their records legit

The question is not whether or not AIs will beat human players in online mario kart games, but rather when. And guess what, they won't even need all these crazy TAS lines to do it, since this is not the point of machine learning based CPUs. You will just think you'll race a smart human player.

Deep learning / neuronal networks are growing at an exponential pace thanks to current hardware and CPU power. Google already announced some time ago that they wanted to apply their AI (Alphago) to starcraft 2 and see if they can defeat the best worldwide players. And if they manage to do it, which I am sure they will, you can consider the job  done for mario kart.

Some of the main challenges (among many others) for that are the same as for any deep learning project:

- find a HUGE training dataset (AFAIK 30 millions of professional go games were used to train Alphago)
- have the computational power to run you AI (1920 CPUs and 280 GPUs for Alphago)
- find people who care about it

Yes you're right, starcraft 2 will probably be a more difficult challenge than mario kart, but mario kart will still remain a difficult task. Keep in mind that the aim of deep learning is to be able to teach an AI varieties of in-game scenarios, some leading to positive outcomes (e.g a SSMT is "positive" because it gives a speed advantage) and some to negative ones (e.g going into a fake item). By increasing considerably the dataset, you might be able to teach the AI things that would make no sense for a standard AI (e.g it might be more interesting to stay second when a blue shell is around, or to avoid being too close to the border of the route if someone is trying to hit you, etc..). But such non standards behaviors are extremely hard to teach because they go against the usual ways of winning the game (most of your games in your training dataset will, of course, teach the AI that going in second position is not "positive" as it does not give victory)

AlexPenev wrote on 01/08/17 at 16:05:20:


Tbh Alex, at first I didn't know if you were being serious or sarcastic ^^
But I think it's good to add some explanations now.

Rhodechill wrote on 01/11/17 at 02:14:14:


Lol, nope they don't!

Rhodechill wrote on 01/11/17 at 02:14:14:


Currently, the CPU's 'AI' of the Mario Kart games aren't smart and they do not learn anything at all to schematize the CPUs of our MK games are just reproducing a pre-programmed code to drive with multi-"checkpoints" on the tracks to go back in the correct way when we put them off the road for example, and it's the same thing for the custom tracks from SMK to MKW and more; programmed checkpoints are created to make them navigate, if we don't do that so the CPUs will drive totally crazy and start driving/looking straight ahead and doing nothing else but hitting the first obstacle they will encounter front of them, in the past we have experienced this on some SMK hack custom tracks.
Currently, the CPU's 'AI' of the Mario Kart games will never beat you (at least if you are a good karter, and here, in this place I assume we are really good!) I mean the CPUs will never reproduce a strat you did before like a trick, a glitch, a bug, an exploit and obviously some of our shortcuts if not all, and the CPU's 'AI' just have easy medium hard pre-programmed brain--(50, 100, 150cc etc)--difficulty for driving/fighting.

This is why the MarI/O program and the TensorKart program are completely different and they are using what is really close to the best AI actually for our MK games, the CPUs will learn and so it can adjust its trajectory to continue, generation after generation, to beat the previous performance, Alex and ROM also explained some points too with StarCraft II, but I try to look at the Machine Learning for /vg regarding the racing/karting category only, but everything is welcome of course.

ROM wrote on 01/10/17 at 09:08:07:


About online modes and GP I 100% agree with you, but as I said in my first post-- it's about TT; and my answer is why not! ^^ Personnally I really need to see this happening to be sure you're right; a CPU's AI in Time Trial defeating/beating the strongest MK64 WR or reaching the last Myth+ time in SMK for example , so you right that it's all about finding people who care about it first, and also findind very good teachers/programmers because, like with a child, we will need at some point to teach the CPU's AI if we really want it learns and then evolve by itself better, by teaching every kind of boost and strat that we know so the machine learning will progress generation after generation until reaching the ultimate best time possible, THE GOAL, and become a true machine's AI. Yup a machine learning without some teaching will never find by itself some advanced strategy of driving like PRBing in MKDS, flat boosting in SMK etc

An exemple with SMW, the machine learning just started with how to control Mario and after 33 generations of evolution was able to complete a level: https://www.youtube.com/watch?v=qv6UVOQ0F44