AI Destroys NES Tetris On A REAL Nintendo! 

Both funny AND true.

Next step for sure!

I talked about this with a lot of people on some of the tetris discords. I'm hoping to work on a reinforcement learning approach that can come up with "unique" strategies on its own, with a goal of maximizing score. To make it useful to human players, I probably need to restrict it to "human" moves. Right now due to the latency issues it probably is doing that by default anyway, but it's still a long term issue. I also am not sure how to best help humans interpret it - is footage of it playing enough, or is it worth giving a guide to what it is "thinking" somehow? Maybe a tetris analyzer where you can upload a board and a next piece and it tells you the best move it sees? I dunno.

@@CodingLikeMad Humans now in tetris can play up to level 38 and higher, killscreen is already easu for them. Ai playing tetris would just be like chess engines which woulf be cool.

Working on it! It'll get there soon!

That's an interesting thought - I hadn't considered that it would make a more efficient method of play by doing that. Score maximizing is definitely the next thing up for sure though, so it'll be interesting to see if that accidently increases survival too! Glad you enjoyed.

That's the hope for me too! That level 29 is quite a leap though. I'll need to replace the whole capture system at least :/

@@CodingLikeMad good luck with it :) it would also be cool to see it go against a human or maybe itself. Just tossing some ideas for you :)

My friend keeps asking me to setup up to play them online. That actually would be pretty easy to do, although I'm sure my account on PPT would be banned very fast if I tried. ;)

Thanks so much, I appreciate it :) I think everyone wants to see the max out, so that's definitely coming. Next up from my side I'd love to be trying out a reinforcement learning approach as well. That's going to be needed for a lot of my projects, so better get started :)

I'm glad you enjoyed it, I've been hanging around the tetris discords trying to figure out how to keep the video appealing to all audiences, from Tetris Grand Masters to people who remember it from childhood :) Regarding your point, I think you could probably display some kind of scores from the trained model for different moves in principle. Sortof a ... tetris analysis engine. Like a "What would TetrisBot do?" type thing. Only problem is how to package something like that(like a website or app or what), the AI itself could just be ripped out of my current engine.

@@CodingLikeMad Are you in our "The Absolute PLUS" Discord? It's linked on the sidebar of the subreddit. I'll take a look at your longer video when I get the time to really absorb the info.

I just joined actually, I'll make a nuisance of myself there soon. Actually, someone asked me if I would port my bot for that game, so probably a good place to be anyway.

Thanks, I'm super proud of this project and so glad you enjoyed it :)

I've been trying to figure where the traffic was coming from all morning, thanks so much for letting me know. I knew aGameScout had looked at it, I didn't realize they plugged it though. I'll have to thank them.

Stay tuned, I'm starting work on that now :) Thanks so much.

Did you check out the linked video below? I talk about the architecture a bit more there. Balancing a general and technical audience is hard :/

Oof, I never noticed that! Poor Tetris bot :(

In principle I should be able to get past 29 with a faster capture card. I think I have about 1 or 2 frames to grab the block shape, and that is doable with good hardware. The AI itself averages only 2 frames, and that can speed up if I need it to. If I can get it fast enough, I could even start doing tspins!

This AI is still too simple in my eyes - stay tuned, the next couple are much cooler :)

Thanks so much :)

@@CodingLikeMad oh just one thing. You should make the ai read the next box so that it's not slow on 29

@@dogenestris I actually do read the next box right now, I just don't pre-compute the move for a single next piece, I use the next two pieces (so the one on screen, and the one in the next box). I could drop to just one piece on 29 though, that would help a bit. A bigger issue is knowing when the next piece is actually on screen though. That type of timing is done very poorly right now, and it wastes a lot of time.

I've got a new AI trained, new video to be released shortly. It's play should make you MUCH less sad :)

Hi Mike, Sorry for not replying sooner, I wanted to give you comment the full attention it deserves! With regards to multi-threading - the original code (as you can see in my bonus video) is shattered into multiple pieces. The code responsible for reading from the camera, and the code responsible for interpreting the output are in different executables. I have a new version of the code that is pure python (and soon will be python+C most likely). This version I hope to introduce some multithreading to, perhaps with a seperate thread responsible for reading the camera. In principle there is enough time to pre-compute all 7 possible "next" piece shapes, however the current system doesn't know the "time" accurately enough to take advantage of this. I hope that by multithreading and properly moving the frame analysis into the same thread that I can close this gap substantially, perhaps enough that I will know if I have spare time to work with between drops. My next video on this is going to use the same features but this time for score optimization (I never listed them, I should include them in the next bonus video explicitly, although I'm going to add to the list so I need to mention it there anyway), but they aren't that different then the ones meatfighter uses - actually, I think they are a strict subset, although I developed my code entirely before seeing his(or anyone elses) work on this. Hoping to hit a world record there, but I need to do the above speed ups because I refuse to continue working with matlab on this. Then I can move on to much more interesting machine learning stuff - I really need a reinforcement learning approach for this, and I'm planning on applying it to Dr. Mario as well :) Hopefully those will match your interest. And while I had thought about your ideas on multithreading already, I wouldn't call them obvious. I'd love to hear more ideas on the next video, since you're certainly at the edge of the things I'm thinking about for sure :)

I know, it's heart breaking, but the good news is that the newer bot plays much better :) I need an update video so bad for this :/

For sure, level 19 start and score optimizations are a big deal. I think, to be honest, that beating 29 is not that important for challenging the leaderboards since it's so difficult to do practically anyway. I'm getting way too excited about the next project though, we'll see if I can force myself to finish my minesweeper video before I start the next teris one.

@@CodingLikeMad Even the best human players struggle to score many tetrises past level 29, so the impact on score of being able to survive the killscreen is minimal. What you need to consistently get high scores is efficient play on level 19 - as many line clears as possible need to be in the form of tetrises.

@@thomasdalton1508 Interesting... that's very cool for this bot, as it can survive pretty well in that level range. Only at 29 does it collapse. I'd really like to see it take the WR for score, although obviously it's cheating, not being human and all.

The AI was actually let loose on a simulation of tetris without any gui, and just allowed to play over and over again, trying different strategies until it found one that let it win consistently. I did a few hundred games with each strategy and it progressively improved until it seemed to hit a wall. Still good enough to hit level 29 though :)

Seems like people enjoyed this, so definitely a few more coming in the next few months :)

I'm most interested in an AI figuring out strategy for itself, but the current version is dumb enough that it's not going to do that. For me, advanced AIs are ones that do things humans dont think of and cant teach. That's my real goal for this, but that's a future video :)

That's for sure a plan, people keep asking for a versus mode to fight the bot :p

It does! I look for the combination of the next two pieces that gets the board in the best possible position.

Noone promised pizza! How do these rumors even start? TSpins are incoming if I ever manage to get a faster capture card :)

Yooo free pizza ???

I'd love to say that's a one time thing for me, but i count like 4 times with even trying :/

I mean, in seriousness, it still plays better than all but top players, that's pretty solid. I do confess that tetris won this round though. Guess we need a sequel.

My AI, it's too smart for me!

Me too! The height creates some annoying issues for the graphics processing that I'm embarrassed to say aren't fully solved yet. floating shapes (like you get when you have height>0) can make it confused on where the active piece is if they happen to be in the shape of a normal tetris block. Def. a thing to play with though, especially since I think the high heights and high levels are known to be crazy difficult.

Since I bad mouthed it so hard I'm trying to avoid naming it. But it is a common one. I can provide recommendations for better ones if you are interested though. Consensus among those ive talked to is that the right way to do this is to mod the NES to RGB and then use an OSSC... I'm just not quite willing to spend that much cash yet on this project.

@@CodingLikeMad that captured video actually looked pretty decent. Which is why I was asking. I've looked into some better ones like gv-usb2.

@@MrK- I'm not a huge fan of the gv-usb2, but it's one of the ones in my repertoire(I own 4 or 5 capture cards now). The latency on it isn't great on it, but if you just need capture it's a common one. One of the things I don't like about it is that it doesn't deal with the interlaced video properly. If you look at my level 29 sequence, you can see the capture card I use has this problem as well - notice how the square block when I first hit 29 seems to occupy 6 blocks? It's showing 30 frames per seconds with doubled up vertical resolution instead of 60 frames per second. Unfortunately, every component/RCA capture card system I have has this problem right now :/ If you compare to an emulator, you'll see much cleaner blocks. That's what I really want to see. But, indeed, the price can't be beaten for that card.

@@CodingLikeMad yeah. Atleast its better of a capture card than what I currently have. I dont have the biggest budgets for better ones.

Yeah, for a low end card, I've found them pretty reliable. No linux drivers is the other thing that annoyed me with them. But the price cant be beat.

Training the AI for score is for SURE the next step. All I need to do is change the cost function actually, so it optimizes the linear weights towards 3 and 4 line clears instead of just surviving. Conveniently, because it uses the total score over the lifetime, it'll actually still consider survival, since your score will lower if you die. I probably will also provide some explicit features so it can see it set itself up for a 2 or 3 or 4 line clear. Hopefully that's enough to hit a max out on the score, since it can definitely get to level 29 already.

I wondered if it would work out to prefer tetrises by itself. You think it may need help?

The feature set I use us pretty simple, I think this version would be helped by adding features that "suggest" tetrises at least. That said, a more sophisticated algorithm like a neural net could learn those features on it's own. Project for a future day though.

Crazy how good the human mind is at this. Even professional self driving car systems dont manage amazing frame rates actually, it's awe inspiring.

Up next, stay tuned :)

I'm skeptical. You'll need to play some tetris to prove it.

Next step for sure!

But it's a good bot, it earned its treats!