If you freeze it right at 9:05 you can see the failure. My guess is the algorithm became dependent on only the straight shape to eliminate rows because the tetris (4 rows at a time) was valued 8 times more than a single row. Eventually there was a long enough stretch where no straight shapes appeared so it failed.

It cant perform complex moves like T-spins so thats a limitation and Im not sure what holding a piece is

@@GreerViau In some versions of tetris you can save a piece for later usually put in a box on the left with the word hold on top. Then the player can switch the hold piece with the current one once a move.

If you want to get into academics watch some online lectures like MIT and Stanford. Otherwise just come up with a project idea and go for it, learn as you go

@@GreerViau what did you do?

Assuming that you don't already know programming, learn C or python, two great places to start. Then if you want to hop straight into AI, watch youtube tutorials on AI projects and follow them the same. At the same time slowly read through and compare your learning progress to this roadmap: i.am.ai/roadmap/#note. As you get better you can start doing your own projects with more complex solutions! Hope this helped 1 year later haha

Yo what are the odds man. Thanks, hope you're doing well

The odds might be pretty high, I watch a lot of youtube lol, I've been good though.

Ok. Atleast u tried

I'm in my senior year of Bachelor's in Computer Engineering and I have never tried AI programming. Salute to you bro!

It doesn't automatically switch to the next state, it might seem that way in the sped up clips but when it finds the best move to make it returns a move list that will get it to that position and thats how it makes moves. The system also has a disadvantage compared to human players that it cant perform complex moves like tspins or tucks and stuff like that. I'm also not sure what you mean by select 1 of 4 moves.

@@GreerViau AI creates this move list? Or it only choose desired state? By 4 moves I meant the following. Input of the ai is current state. Output is a move to make in the game: left, right, down or rotate.

@@alexrodichkin7172 well like i mentioned in the video, the ai is responsibe for selecting the best move. A move list is just a set of inputs that will get the piece to the chosen position. Also like i mentioned in the becoming of the video, i first tried what you were suggesting of taking game state and outputing left right up down commands but that didnt work. The reason is for the ai to learn anything it would have to randomly stumble into reward which in tetris is really hard and almost impossible

@@GreerViau That is why it is so interesting. We humans can master it this way. Your ai mastered simplified version of tetris. A lot of information given which player of tetris has to think by itself. Please tell me how far did your ai went without help? I am asking because I failed to teach it some time before with ga.

@@alexrodichkin7172 Using a genetic algorithm it wasn't able to learn anything with the "human" way of playing. Genetic algorithms just arent powerful enough to optimize neural networks for problems of that complexity. Not to mention very slow, training with this style of gameplay took hours to complete 1 generation. I did first attempt a "human" level of play using a Q learning algorithm in python, and that may have worked but reinforcement learning still takes a very long time, I left it running for 4 days and it had barely learned anything. Again the problem with tetris is that if the ai is going to learn how to play with no knowledge of the game, it needs to randomly perform an action that merits reward and in tetris that is very difficult to do. Not to mention even if the ai performs an action that merits reward, in tetris that action may not always merit reward with a different game state. When we play, we already understand the objective of the game so we have an advantage.

"Never thought I'd learn a new thing from youtube" if you really think so i bet you are either an old man discovering youtube or your ego is bigger than this galaxy.

The code is all in the repo

i suspect it won't do so well since it can only look ahead 2 pieces, and i think hatetris requires quite some forward planning for decent scores.

Well you don't need the neural network to make this work, calculating all of the possible moves is just a brute force approach. The neural net comes in when you need to score each move to decide the best one to make. But technically you could make a scoring function without a neural net, which I did at first in the video, but using a neural net was a better approach for the reasons I explained in the video.

Processing

@@GreerViau ok, thank you

ah yes, splashy splash long shapy shapy

Indeed, a tetris-tetris

The calculations get faster when there are less players alive in the population. The population starts out at like 200 or something so it's very slow, but as players loose, there are less players doing calculations so they can do it faster.

i don't think the network works in real-time. it just evaluates every single possible move. it takes longer to make evaluation two moves deep. but that does not makes it harder for the network, it just takes longer to train it.

Even with next piece knowledge, the ai still makes its evaluations as soon as the piece is in play, the only reason it seems to pause is because there are 200 players in the population. So regardless how fast the game gets it will still perform the same. I didn't mention in the video but in those sped up clips, the system is already playing 10x faster than a person but I had to speed up those clips in post, so when it says 10x speed it really is playing 100x faster than a person.