edit: is this your article? thank you for the wonderful post!
> In practice, IL programs direct the agent to construct a pile near the origin that progressively grows taller and wider. That being the case, it can work with a Tetris implementation that defines “row infinity” as a finite row whose index increases as a function of the number of spawns. In such an implementation, the agent emulates a semihard drop with a finite number of soft drops.
> When a newly spawned piece falls, it never gets closer to the floor due to the nature of infinity.
i'm nitpicking, of course. everything described is possible with a sufficiently large board. you'll just have a time constraint to place each block.
I don't think the infinite board is that big of a deal. Turing Machines and Lambda Calculus have no limits on their memory space and we implement analogues of them in our limited memory space.
> On an infinite playfield, tetrominoes spawn at “row infinity” and column zero. When a newly spawned piece falls, it never gets closer to the floor due to the nature of infinity. This means, in finite—though potentially vast—time, the agent can shift the piece into any finite column. And once in position, the agent can hard drop the piece.
gate construction relies on "nudging" a block at a specific row as it is falling. but if blocks start at row infinity, the only way to get them down should be to hard-drop them -- which precludes nudging.
they've defined a new "semi-hard-drop" operation, and introduced it as a "generalization" of legal moves, but it isn't.
For instance: "For instance, if the agent emulates a semihard drop with a finite number of soft drops, then it will let a gravity drop serve as the final soft drop. "
Consider this. Our physical implementations of turing machines work even though we don't have an infinite memory. We do encounter some limitations because of such.
The Tetrino also works. You can create it and run it on a finite board. It would be limited just at our physical machines are limited without access to infinite memory.
Not to mention that you need unbounded memory anyway.
the infinite board solves piece randomization, because you can have a "junkyard" where you dump pieces until you get the one you want. it also solves the time pressure, because blocks spawned at row inifinity will never reach the canvas until you hard-drop them. finally, it prevents rows from clearing because they have infinite width (so the game won't interfere with your construction).
their concept for gates is that if you nudge a block at a certain row as it's falling, it will either shift one column over or be blocked. this page is a good visualization:
https://meatfighter.com/tetromino-computer/inverter-not.html
except... once you hard-drop a block in tetris, you can't nudge it. and if you don't hard-drop it, it will never reach your canvas.
they kind-of hand-wave this in the section "infinite playfield". after describing the mechanics of a hard drop, they say:
> A more generalized version, the semihard drop, attempts to vertically translate the piece from “row infinity” to a specified, finite row.
if the fall timer can't get the block into a finite row, how would a player (without hard-dropping)?
loved reading through this, though. the animations were perfect for describing the mechanism! very high-quality documentation for a fun project.