4:32 song?

the AI drives exactly how I do in racing games, all throttle, no brakes XD

what are your computer specs? its impressive that your computer can run multiple ai simulations at a time!

These are the ultimate KZbin videos. Secretly educating you in some way disguised under a very entertaining video.

What kind of computer can generate that???

I hope so desperately you return to the topic of chaos in TM. Like if you force conservative play (such as a short life with an enormous penalty for falling), and then slowly reduce the time restrictions during training. I imagine it would be possible to find the approximate point where the chaos is amplified beyond what can be reliably controlled.

Considering the backwards run seemed more stable, it would have been interesting to see how that approach worked on the third level. Regarding creativity, maybe you need to teach it impatience by adjusting the reward based on the how well the progress between different iterations of the network improved.

amazing

How do you make the AI interact with Trackmania, getting the information into the AI and the Controls out sort of thing? Is it generic or Trackmania specific?

Love These Videos

This is amazing ❤

bro is so good at trackmania that he had to make an unbeatable AI to beat him

Sir, do you mean to tell me, that you design an AI to be insanely good at extremely aggressive strategies... and then run its deterministic engine, THOUSANDS of times, so that you can figure out what steering angle to hold for a few frames so that it will suddenly work...

When you came up withe the indeterminisic pipe problem, i could bet you would come up with chaos theory. my best guess is the fact that in deterministic games, there are random fluctuations that make any action unrepeatable over infinite time because of wave superopsition leading to superposition of recurring errors adding up.

Can't outsmart Nature. God is it. Lots of scientists about to cope bigly.

Yes. Consciousness is the future, and Western civilization is on the decline. China is rising.

I'm disappointed you didn't use this on a more elaborate map with jumps and having to use tricks that a human can grasp on it's own.

Entertaining and educational 🧡

are these the random tiktok lives i see

Today I learned that trackmania car eats carrots XD

Time resolution is surely a limiting factor, eapecially for consistency

how about trying a RNN , the specific feedback from the last decision might help bring up the learning gradient I dont think i conveyed my idea properly but i have a paper publish on this. Do contact me if u need to use it.

Ai with cs match?

@wirtualtv @nadeo

5:29 song?? is it yours or something by lapfox or patrcia traxxon

What a sad ai.

This is insane. I'd love to read the code. Congrats.

Then imagine a quantum computer being able to simulate something of this scale in a few seconds instead of years lol

You must add braking.

I remember when this came out and now it’s 7 months old? What happened to time 😭

Wish I could give this video 10 likes

Hi. I have a quick question. When you say 100 hours of training does that mean like 100 car training for 1 hour or 100 hours in real time regardless of number of cars?

I think the graph shows that the ai is facing problems. Not anything really big but theoretically, each generation should be faster and depict a straight line graph with a strong negative correlation. Isntead, it depicts a consistent curve and at times shows no improvement with the more blocky portions. I‘m probably misinterpreting the graph though. I did also want to mention that, theoretically, it‘s very unlikely that a person can successfully make an ai smarter than the creator. It has been done but a single person can only teach what they know and machine learning is like teaching a pet where they don’t truly learn the skill, they just learn what to do for a reward. AI isn’t my area of expertise so I‘m probably wrong in some way. I just like maths and social science :)

This is beautiful, really makes me want to learn how to code again

From my knowledge the input to the NN isn't nodes with "Distance to centerline" or "Next Curve Distance" it should be the pixels on the screen, the AI then take this input and extract the feature in a latent layer (if they are relevant) and the output of the AI is the input to the car controller (gas, breaks, steer left or steer right). The idea behind it is that some figures might not be reachable by the agent (if you want to make a fair agent that plays "as" a human), and some of the figures might not even be relevant to the agent. The actual input to these reinforcement learning is the pixels on the screen, from those the AI "learns" what works best in these situation, and then generalize it to "situation that looks like this" so if the AI sees pixels that look like a curve is ahead, it would learn to slow down if that feature (curve ahead) is relevant to its performance. About the AI "loves crashing into walls" you should define a new reward function to maximize, look at say 2 minutes period, make it max distance (on track) from start, and don't even look at "race finish" at all, so for example if a race should take 3 minutes if played perfectly, each training session would be 2 minutes and the best score would be the one that reached the furthest (that's how I'd train agent for a racing game). About the generalization, the "pixels as input" approach would be much better because like NN can be train to recognize things, it can recognize map's curvature, it's own orientation, and things that u wouldn't even think about adding as data, things that you wouldn't even think relevant. Also, a minimap is included so if it "decides" during training that the minimap data is relevant it will get the data as a feature in a latent layer. Also, pixels as input approach can be a general AI to any game, the only thing that you would change is the reward function to maximize.

If yoi use math instead of tentatives that can be more fast. If you put the AI to scan all the track just one time and it made de curves with mathematic, make the processes more effective and the AI can do the perfect race with less time

c est quoi cet accent de merde

No more vids :/

how does the ai realise that the reward is actually a reward and not just a random number? Why does it want to get a higher reward

So technically I could beat the ai if I tried long enough

Ça se voit t français

bro got a deep rl phd student and a roboticist, because of a trackmania video.

Yo bro why do the F1s have real car rims and tires 😂??

This is not AI. This is simple Bruteforcing

*I am pretty sure, according to chaos theory, outcomes are completely predictable. It's just that U will have to be too accurate and will have to gues every single microscopic event happening during the run to find out the outcome of chaos.*

isn't there more observations to be given regarding the track we can see using eyes to the model? does it see a t section or a left turn or is it driving as blindfolded and remembers the map from memory/past experience?

this is probably top 5 videos ive ever seen. ive never been so interested in something while watching. gg's dude

I wish you linked the source code

I would add a timer and it gets more points The faster it finishes

When your only worthy opponent to race is yourself