Last night I was laying in bed wondering if anyone had done this and now, after playing a bit of minesweeper, I remembered that I wanted to look it up. Great stuff.
@algola7024
3 жыл бұрын
Yo you're so awesome 😎 Also that large board was oddly satisfying
@CodingLikeMad
3 жыл бұрын
Thanks! Somone asked me to just a bunch of those back to back :p
@Failstreamtwitch
3 жыл бұрын
Cool vid, Looking forward to the next one!
@CodingLikeMad
3 жыл бұрын
Thanks, I'm really pleased with how it turned out :)
@foiledthrice
3 жыл бұрын
POG new AI finished
@CodingLikeMad
3 жыл бұрын
I love how every time I say "a couple weeks" it turns into a couple months. That's the best.... :P
@jeremyhastings4756
3 жыл бұрын
So cool!
@mohamadarsyakaukabi3317
Жыл бұрын
How do you calculate the Avg human win rate?
@kutubahemad7937
Жыл бұрын
Why u are not making video brother
@hanskraut2018
Жыл бұрын
add more machine lerning and neural network techniques and make it generalize to other games. Find the simplest game in terms of simplicity for the strenght of the neural network not yourself as a human and then overcome what it cant do but only do the most simple things first. Add internal repersenations and diverent layers for different time horizons and a way say previous weights / brain states and intermesh them or "summarize or hybernate them to give more space to the current task etc. Also let it write internal represenations and let it edit a small part of the neural network (not all or disaser) like PFC in humans selfregulation speaking to yourself executive functions. Also longterm memory stuff. Just build on top of things. Would be fun to see how it evolves with fun challenges pictures and short conside vids like this. Can use stuff already existing. Try to solve a small thing that has never been done before (and if it did then do it wiht 100000000x less space/compute) :) fun
@johntanchongmin
2 жыл бұрын
Never knew a neural network could learn so well for minesweeper. How long did it take to train it?
@CodingLikeMad
2 жыл бұрын
I think the time lapses are real time, so something under a day? It's a pretty simple model, since it is supervised learning, so its just a time question really.
@CodyGall
11 ай бұрын
Can you make an AI that feeds the hungry children of the world?
@hydrantdude504
3 жыл бұрын
how'd you get the "average human" win percentages?
@CodingLikeMad
3 жыл бұрын
Those are from the World Of Minesweeper all time statistics board. For the easy mode, I had to search a whole bunch because my easy mode is slightly different than the one on their website, so the stats on that one might be a touch off.
@isamiwind438
3 жыл бұрын
@@CodingLikeMad those human win rates aren't exactly right because players on WoM usually go for speed which can mean taking some guesses that can be faster or maybe even accidentally misclicking. I don't have much data but I'm pretty sure that for expert a human can reach at least 20% win rate. I don't know about intermediate nor that beginner (8x8 10 mines). Though on 9x9 10 mines (beginner from WoM) win rate of ≈80%-90% is also possible. Not sure how this algorithm works but I want to mention one more thing. While mine/safe probability are important another important thing is progress. That is some cells are better to guess because they give more probability of further progress.
@CodingLikeMad
3 жыл бұрын
@@isamiwind438 For sure they aren't really comparable, it's intended to give a ruler to give an idea how good/bad the bot is. After all - people playing on world of minesweeper aren't exactly of average skill either. It's a very biased sample from a lot of lenses. I hoped that by saying "average" I would avoid giving too warped a view on it though - I know some others have claimed to have the best possible AI, and this certainly is much worse than the best human in terms of accuracy. I should note that it is also still pretty undertrained in those games, I only pushed it for ~24 hours of training, 72 I think at most, and I've since sped things up quite a bit. Someone provided some numbers for algorithmic solvers in one of the other posts, so I'm hoping to compare to those once I've trained it heavily, and really see how it measures up. Your point with higher probability of further progress is a very interesting one as well actually. I've been trying to figure out how to make more progress on that front - there are AIs which can make that kind of calculation, but combining it with this might be pretty challenging. We'll see how much interest there is in the community, since it is a pretty big project. Thanks for the feedback, hopefully next time I show it off it can be closer to those numbers :)
@isamiwind438
3 жыл бұрын
@@CodingLikeMad Good luck
@zc2012
2 жыл бұрын
the win rates of AI are not very impressive
@CodingLikeMad
2 жыл бұрын
Compared to what? I'm curious if you have a reference for a higher performance. The mechanism chosen for it is likely to under perform relative to brute force state search, which is what the higher end AIs do for minesweeper once it is a moderately constrained problem, so I wouldn't think this is the highest available numbers, but I do provide the relative win rate compared to humans - which it blows out of the water.
@zc2012
2 жыл бұрын
@@CodingLikeMad Sorry, I happened to watch another video whose win rate looked higher than yours, but I believe he used a different mechanism. In case you want to take a look, here is the address: kzitem.info/news/bejne/1qyIt4mNsYF5pKQ
@CodingLikeMad
2 жыл бұрын
@@zc2012 Oh, that's a very nice video, it shows a nice comparison of academic attempts at this too. My AI is very much not tuned to be the best possible, so I'm ok with this performance. After all - I would have bothered training it to extremes, or for that matter trying a method that could inherintly capture long distance probability effects if I had intended to do better :) It was mostly intended as a demonstration of a neural network working on a problem that it isn't very good at out of the box - and beating human play when doing it. The tricks that he used - some of them my AI has figured out on it's own, and I find the idea of a machine teaching itself to play very interesting, even though it isn't the best possible solution. That said, I AM trying to achieve that with tetris. I've got two tetris videos out right now, and I'm hoping to get one more that really shows the strength on real hardware before Greg manages to break a real console and not just an emulator :/ Thanks for your link!
Пікірлер: 25