I'm a huge fan of Tetris and have played them all. I think this version is the best one. Not only because of the drama behind it, but two player mode is actually pretty fun.
Recently got Tetris Grand Master 4 and played it with an arcade stick insanely satisfying to play that way and with that ruleset and kinda ruined other versions/input methods for me.
I was expecting something like DQN, but what I actually saw was a new approach, so it was fascinating. Usually when you're making small AI demos and doing hands-on exercises, you work with Tetris a lot.
In NES Tetris, if the input is the same, the result is the same, so you can store all the inputs and reproduce specific moments. The state becomes like a graph, which allows for fuzzing testing. It's interesting
Seems to be influenced by the pieces on the board. Bill talks about it a little in the article. You do seem to get more | shaped pieces when you leave those spaces open on the board.
Nope, piece selection in the unmodified NEStris uses only the player input to generate pieces and it also starts from the same seed, so frame-perfect inputs after loading the game will result in the exact same pieces. The first perfect level 18 run (only tetrises to score) was dismissed by community because it was made by replicating a TAS run with a suitable piece sequence.
> You do seem to get more | shaped pieces when you leave those spaces open on the board.
I don't think this is correct, nor that it can be evinced from the article. What it does say is that the sequences that led them to achieve their target show a higher incidence of I shapes. This is because all the ones that show less I shapes have been "pruned away" by the cost function, which favors I shapes.
This has some relationship with the anthropic principle: isn't it strange that, of all the possible universes, we ended up in the one that seems fine-tuned exactly for life as we know it?
Their strategy optimized for the I pieces, so it should be no surprise that winning runs were ones that had a higher incidence of those pieces. Runs with fewer would not have survived under that strategy.
The best versions of Tetris are random in sets of 7. There's a "bag" of all 7 shapes and the next piece is picked from that bag until the bag is empty and then it refills.
You're talking about piece, right? I know that it's actually determined by the execution state per frame and the button inputs [1]. I looked into Tetris AI before because I found it fascinating, but I might be mistaken.
https://tetris.wiki/Tetris_(NES,_Tengen)
In NES Tetris, if the input is the same, the result is the same, so you can store all the inputs and reproduce specific moments. The state becomes like a graph, which allows for fuzzing testing. It's interesting
I don't think this is correct, nor that it can be evinced from the article. What it does say is that the sequences that led them to achieve their target show a higher incidence of I shapes. This is because all the ones that show less I shapes have been "pruned away" by the cost function, which favors I shapes.
This has some relationship with the anthropic principle: isn't it strange that, of all the possible universes, we ended up in the one that seems fine-tuned exactly for life as we know it?
[1]https://tetrissuomi.wordpress.com/wp-content/uploads/2020/04...