MineSweeper Big Board AI Solves [60 Minutes]
Вставка
- Опубліковано 25 лют 2021
- By popular request: 1 hour of my minesweeper AI obliterating a full screen board :)
I created an AI that demolishes Minesweeper in real time using Neural Networks. The AI taught itself how to play through trial and error, taking many thousands of games to master minesweeper - but now it's better than the typical human by a wide margin! This project originally dates from April of 2018, but it's taken forever to get out, so I'm proud to finally be done!
If you want to watch me make this stuff live, checkout my twitch:
/ spshkyros
Check out my other social media if you want to follow my work:
Twitter: / codinglikemad
Reddit: / codinglikemad
The original minesweeper AI video is here:
• AI Neural Network Beat...
If you want more details on the algorithm, I posted an unlisted video with more details here:
• AI Neural Network Beat... - Наука та технологія
The original AI video can be found here: ua-cam.com/video/lN-Pq1GoIO0/v-deo.html
Not sure why this is so hypnotic, but I can't stop watching it, glad some of you feel the same :)
It's satisfying and so cool at the same time!
I've been struggling with getting other work done, because I want to just watch the AI work. :/
I was hoping for a 99 hour version. its soooo good
People seemed to want 10 hours, 99 seems a bit... extreme. I'll see how this one does though, and adjust future long plays accordingly :P
Need a running scoreboard up in the top bar (both the white and the gray). "Games Played: Won: Lost: % Improvement: " Positive improvement in green with an up-arrow and negative improvement in red with a down-arrow. I wouldn't count one-square fails as that just means their first square had a bomb.
I think a running score board is a really good idea! Next time I do a time lapse, I'm absolutely including one. I have a rough one I used in my last minesweeper video, but it's not clean looking, and I think that's needed. One square fails are interesting because human stats count them, so it's sortof apples to apples... maybe the next tetris video is the next time I get to apply this though. Thanks for the feedback :)
...... and afterwards the machines gained consciousness and made endless minefields in the war against their human overlords!
You should rewrite the minesweeper game so that when it reaches a certain % complete the board expands by 2x the volume. Then train an AI against it until the AI becomes sentient.
"With KyrosNet Activated humankind will finally reach its potential"
-Some movie, probably...
My concern is less that humankind will reach it's potential, but that future historians will ask why noone stopped me :/
Kyros Net has a nice ring to it
@@algola7024 Don't tempt me. ... Who am I kidding, TEMPT AWAY! KyrosNet DOES have an amazing ring to it, doesn't it?
Most importantly what's the win loss ratio?
I counted by hand so I could reply :P 99/500 games is just under 20%. It's the most heavily trained AI I've made by far actually. It crashed my computer once during training :P
@@CodingLikeMad So much training but sadly your AI still does simple blunders. One example from the video would be at 57:07. That field has many things that could be logically and definitely solved but your AI takes a guess. Did it think that the cell was THAT safe? Or it simply didn't see any "likely safe" cells? (there are definitely safe cells as I said)
is this a "zero" learning algorithm?
I'm not sure what you mean by a zero learning algorithm, but this uses traditional supervised learning. It is using the input board that is visible to predict the mine positions, and then recording every time a game ends to give a new data point to learn from.
To be clear, I'm using a convolutoonal neural network to do the calculations. Most of my newer AIs use reinforcement learning, but this seemed so straight forward with classical supervised learning that I kept it that way.