Video

While I am rooting for you, I secretly hope humans still come out ahead. Their determination and skill truly show how some people are on a different level from the rest

Did not expect to see the best ever visualisation of local vs. global optimum and gradient descent 3:20

This is super cool! Thanks for sharing.

Imo its not a traxkmania "playing" AI, until it (simulates) inputs eith a keyboard/gamepad and reads the screen and got no emgine access. You make it out like there is a gradiant between TAS and normal runs, but there is not. Its either one or the other

Imagine we actual made a fly brain play this game and give it rewards/punishments for good or bad moves 😂😂😂😂

14:20 did not win on strat.

8:10 this took half a year to make??

6:05 with triple hardware can we get that to 60 fps

5:30 Please take this AI and put it into another game, something equally simple, though good. I'm not going to defend that. I'm not saying anything. Put it in another game and I want to see it learn another game and then put it into another game and then bring it back to the first game, you know?

3:30 Beautiful display, thank you. I'm wondering why, then, we do not induce reward for a bored, uh, jovial, experimentation and expediting into the unknown. Why not try something new, dear AI? Here's a cookie.

please stop making skynet

Isnt this just a TAS?

This video shows one of the biggest drawbacks of training AI this way: The AI does not understand the game in general, it only understands a single level. You see, a human driver who has played only one level in his life, and has become really good at that level, can play pretty much any other level from the very first moment. He may not be good at that level at first, but he will be able to play it at a competitive level once he knows the track because he understands the game itself, the controls and the mechanics. The AI, on the other hand, has to relearn all of that for each new level, it cannot use the knowledge it gained from playing one level and transfer it to another. And that's one of the things that still separates AI from true intelligence. The ability to apply knowledge gained in one situation to master a completely different situation is a key element of intelligence. Even higher class animals can do this (and by higher class I mean birds and mammals, not just smart animals like apes or dolphins). This is also where understanding the past and the future comes into play (transferring knowledge from past events to possible future events and thus making predictions about the future and reacting to those predictions before anything has happened). AI is not currently capable of understanding anything in a larger context. It can only analyze the situation directly in front of it, understand simple if-then logic, and predict the next step to take. Even ChatGPT works this way: It only predicts the next word based on a given context and what word it has printed so far; it has no real understanding of what it is doing at that moment.

You should go back and show it some of the short cuts and tell it the reason for why it improves times and then let it add checking for similar stuff with a low weight (like ever thousandth run it will purposely try something new for five or six runs if it diesnt find an improvementit wont try that specificthing in that spot again and the thousand runs start over again. Of course this only happens after it has completedthe race like a hundredtimes successfully). That way it mostly doesn't check but every so often it will give it a try and if it finds an improvement quickly then it will keep trying and if not it will stop trying. Tgat way it can look for shortcuts but won't obsess over finding them.

"Stay with me..." No.

Ya but records are for humans.

Let's see your AI tackle Deep Dip 2

give me the code, I'm 100% gonna try it on some random roblox players.

If you so smart, why dont you create game and make the AI buy them you muppets!

Could you make the AI learn from player ghosts? It could learn to determine which players to trust and which not to - and discover shortcuts.

Does the agent understand it's place within the track? Ie is it keeping record of which lap it is on, and where it is on that lap?

Shouldn't the agent already know how to turn, from the get-go, no matter which track it is on?

You did not trained AI. AI does not exist, stop using words you do not understand. People will start to think YOU do not have "I".

awesome!!!!! Thank you to all the human players who thought outside the box. and congratulations on your great training model. now lets solve the energy crisis and food in the world...lol one day maybe

Writing a paper on this I hope?

Before seeing the end of the video, can you make the gamma function a variable that evolves with each iteration, so the AI can try out long term strategies as well as short term?

"Evolution is smarter than you." Genetic algorithms and neural networks are so cool. Leveraging evolution this way is so evil.

Try feeding it the z buffer rather than the light

Some nice visualization in this video!

make it drive oach

The real test is to tell the AI, once it has beaten the human score.. to beat its own score. Then see if it can work out some crazy stuff. Like can it beat its own best score by .06?

A pathfinding algorithm between the different checkpoints might go hard.

You're an incredibly talented programmer, content creator and video editor. God damn.

Is anybody scared this could ruin TMX leaderboards?

So now what are you going to do next? Beat your AI's time?

human assisted speedruns

Nations?

a few years down and it's gonna be 1000-1

In AI, nothing works better than getting unimaginable amounts of data into an architecture that can absorb at all.

list of all techniques used in a game

most interesting video. thank you! really enjoyed that

What is the problem with A01 map?

I see a paper in HCI and gaming coming soon :D

One thing I've always thought about these trackmania ais is that they play based off of instinct and with no theoretical basing. With purely game based training it should be near impossible for an ai to get a deep understanding of the mechanics of the game, and therefore it will always lack some level of generalized skill only obtainable from said understanding. One reason why i think this is that humans are like this as well, if we spend a ton of time improving one thing we get reqlly good at it even if we have no true understanding just intuition and muscle memory, but as soon as we gain that hidden knowledge we can incorporate into our intuition and reason with it to make more informed decisions. With that i suggest that one possible way to improve a trackmania ai's ability to generalize its skills is to give it access to techniques and knowledge that we've refined over the years.

In my thesis I worked with a paper called PIFU, it's used for 3d reconstruction. Basically you take images, run them through a conv. network and then sample the features at a XYZ coordinate. Do you think it's possible to give each block an ID and sample a 3D grid with 3d convolutions for the current position? That could give the AI a complete overview of the map in 3D space.

Shortcuts? naaah, i will take regular path :DDD

Wirtual really said "I make 4k a month being the best at this Ill give you a c note if you can beat a machine thats better than me"

That's yet another amazing step forwards from the previous - are there no limits? Really looking forward to your code too!

I heard there is a method that GPT trains another AI. I don't know if it's possible to do in this case but it would be fun to see an AI train AI from scratch

26:42 the time is 53 64 i saw though it