AI Learns LASER HOCKEY 

Yeah simple games do seem to give a really nice smooth improvement as the AI trains!

I think that's because that angle makes it really easy to squeeze the puck between the bottom wall and the paddle to get high-speed shots consistently. I know because I used to use that against my brother lol

Wii Tank game would be amazing to see, but definitely a real challenge! Might have to leave my pc training for about a week to beat that one haha

Honestly I have no idea! Not sure if they would just become like gods and no one could score, or if they would get really good at directing their shots so the other player couldn't block them.

​@@aitangoWell, there's only one way to find out!

The way the red paddle tracks at high speed looks so goofy

Yeah that would've been entertaining, I might try that on some other videos. Is it 14 points to get the platinum medal? I might've gotten it while training a couple of times since it seemed to be close in most of the games I watched. Perhaps I should've let it train for longer!

@@aitango According to Google it's just 14 points. I wasn't sure if maybe it would penalize you if the CPU opponent scored, but I guess not. I wonder if it's possible to get scored on 14 times, and score 14 goals in the same game, and tie and still get a platinum medal. Maybe your AI is the bot for the job?

Really glad you think so, I’ll keep at it!

Probably got nervous it would score an own goal

Glad to hear you like it! Yeah I think it just gives that extra flexibility, and also bouncing it off the wall just seems better in almost every case than just hitting it straight

I thought you knew about my channel! That’s cool if you came across it anyway though!

Thanks, great to have you around! Probably won't make a video on it, but here it is! This is run on my own PC rather than cloud compute, but was very expensive to initially buy. The GPU is an RTX4090, the CPU is a i9-13900k, and I have 64gb of ddr5 5600mhz ram. All of the mentioned parts massively speed up training. GPU allows you to do the learning (backpropagation) faster, the CPU allows me to run many environments in parallel very fast, and the algorithm im using requires something called an Experience Replay Buffer, which can use huge amounts of RAM (often far more than 64gb!)

Yeah I briefly checked over them but was still a bit asleep when reviewing them haha

I’d also strongly recommend moving the hours trained display thingy away from the top row of subtitles. They’re the same color and everything, so often when reading subtitles, it’s quite confusing.

Thanks! I learned a little coding in school when I was about 14, then I got really into making games so spend hours learning to code that way! I have mostly stuck to Python, but also done a couple of others here and there

@@aitango thanks for ur comment, ill look around!

Yeah it really mastered that shot! I would absolutely love to do COD Zombies, definitely be something I will at least give a try…

@@aitango Hey how could I learn to build an AI like this? I follow the competitive cod zombies community and I'm curious about the potential use for AI clearing confusion with certain strategies (particularly on Shi no Numa). I've drafted a list of variables an AI might want/need and have decent game knowledge regarding BO1-BO3. If it might matter, while frames per second is known to affect movement mechanics and input frames in games like COD, the "server side" (which applies even to solo mode) runs at 20 Hz, so it might not be as resource intensive as it might otherwise seem just having the AI read variables at that rate. Also, low FPS can even be advantageous for movement in certain situations over high FPS. I don't know much about AI so I don't know how naive I might be about this 😅 but I would be willing to train it a long time if there is real potential for it to discover novel ideas.

Thanks for the sub, glad you liked it! The AI and the CPUs work very differently, as the CPUs use quite traditional techniques (this game came out almost 20 years ago), whereas my AI is a self-taught neural network

mhmm

I'm very glad you enjoy them!

I actually tried doing Wii Tanks a while back but had some trouble with it, but its something I want to come back to because I really loved Wii Tanks :)

@@aitango Awesome to hear, i hope to see that video eventually.

The technique to learn is Reinforcement Learning, although it can take a while to learn if you are new to it! I don’t currently, but may create one soon! Also, I might currently be planning another video on NSMBWii haha

Rainbow is pretty popular, only algorithms I see more often are probably DQN and ppo. It’s a very strong algorithm, but fairly complex to implement. Great for learning rl though!

Yeah it really learns to go for wall shots the cpu just can’t seem to handle

That would be interesting, it would probably take a little while to adjust since it appears to really like using the panels at certain angles, but may find it easier in the end!

@@aitango Woohoo, more content!

The Wii CPU is just following an algorithm, or set of instruction that do not change as it plays, hence it has no capability to learn. This AI however is able to modify itself to improve! Maybe I could play against the AI, I'm not sure if I'll do it for laser hockey for Mario Kart its a bit of a dream of mine to eventually do, and get the AI to beat me!

Yeah I would love to do most things on Wii Sports, but sadly the defining the actions is quite difficult since they all use motion controls! Plus at the moment I actually don't have access to the controller's accelerometer, so can't perform shaking motions :(

@@aitango cool and i did not actually expect you to respond

I’ll accept that as a challenge

What do you mean by adding more CPUs? Also I’ve recently been looking into using bigger neural networks for my next videos! Sadly it does often slow training down even if it leads to a better result

@@aitango I think they mean running multiple instances of it.

@@aitango multiple bots against the ai

The video is very old and probably much worse in terms of editing and stuff, but I actually have a video on pong!

I actually looked at doing this, but its surprisingly difficult! There are lots of different actions which really slows down training, but the biggest problem was that the AI bullets are really small, and the AI already has a very low resolution image to work with. Its definitely a problem I wanna solve though!

?