swng

Пікірлер

@gavrolux409 5 сағат бұрын

socool

@xynix1549 Күн бұрын

i love 20tsd

@algebruh4185 4 күн бұрын

This is what it looks like when my replay desyncs.

@algebruh4185 4 күн бұрын

The noticeable speedup when you convert into ST from LST is criminal 💀

@algebruh4185 7 күн бұрын

first underwater zenith speedrun ever

@hepta5040 8 күн бұрын

Your vision is crazy good. Very unfortunate :(

@algebruh4185 8 күн бұрын

this person does slow cheese at the airport point and laugh

@algebruh4185 10 күн бұрын

i think the red circle improves the thumbnail

@swng314 8 күн бұрын

I made video, then made thumbnail, then looked at it and decided the pace calculator wasn't clear enough, so I remade thumbnail with a red circle lol

@awz2142 14 күн бұрын

Now do mechanical 10 pc

@Yukata779 15 күн бұрын

very nice keyboard

@snowman99tetris 16 күн бұрын

Huh

@파백 16 күн бұрын

perfect

@notmelenda 25 күн бұрын

Pain.

@illegal_melon 25 күн бұрын

sorry idk what was happening the sfx got me

@algebruh4185 Ай бұрын

This is still NN, correct?

@swng314 Ай бұрын

Yes it's the same general alg I've been using recently with the NN solver at its core

@urugiro8950 Ай бұрын

Will it be better if the biltz was the same as the new tetra league getting b2b when pc ? And just spamming tetris sdpc dpc run ? Without breaking b2b will it be efficient ?

@swng314 Ай бұрын

Well right now the bonus for b2b in blitz is just 1.5x (TSD goes from 1200 to 1800) so the penalty for breaking b2b isn't actually that big a deal, so it's always worth breaking b2b for a 4L PC if necessary. If the scoring system was revamped to incentivize holding a chain to the point where skim PCs just aren't worth taking, things would change a lot... Quad PCs are RNG. Parity rules make consecutive 4L quad PCs very RNG, although maybe allspin helps this a bit - not sure. SDPC DPC is very good if you hold b2b, but the RNG on SDPC holding b2b isn't that high. There are some 14L loops that preserve b2b and have high quad PC chance... I think most notably STD/Substitute - e.g. kzbin.info/www/bejne/aHm8doapfdRoprs - but I don't actually know how consistent this is, the branching paths are pretty complicated (this was a set seed) I'm thinking the optimal method would be to spam tspins/quads in a method that allows you to take a b2b PC if queue allows at regular points.

@파백 Ай бұрын

now we can wait for 3000000

@BeCrazy_ Ай бұрын

a thousand more score, so worthit

@Joanyan Ай бұрын

only 100k more score that rtxile 😨

@BeCrazy_ Ай бұрын

what the kb review The Keyboard Artist.

@F1e308 Ай бұрын

I want to see how many points you can get with looped queue of “40L 13PC” at 5pps

@swng314 Ай бұрын

I think I need a particular seed to run with. I can set seed for a TAS run, I don't know how to set queue in blitz.

@F1e308 Ай бұрын

@@swng314 then where is the Quad PC every 4 lines?

@swng314 Ай бұрын

@@F1e308 Consecutive quad PCs require the right RNG allowing even Ts to be accessible every 10 pieces. Is there a seed that allows that ? idk

@megamaz108 Ай бұрын

why does the bot stop and think

@swng314 Ай бұрын

I've set it to match the pps of the current human blitz WR. It's not allowed to place the piece until the human places the piece.

@megamaz108 Ай бұрын

that's pretty crazy wtf but why? wouldn't it be more optimized if you... didn't?

@swng314 Ай бұрын

@@megamaz108 It can pretty much go at whatever speed I set it to. I just decided to use the speed of the current human WR as a baseline so you can compare the efficiency with the WR, I thought that was a reasonable choice

@megamaz108 Ай бұрын

@@swng314 I see. It's just that generally TASes ignore human limits and go for max score, but if this was your intention that makes perfect sense. Very cool stuff! Also, how did you make this? Is it replay editing, slowdown tools, or something else?

@swng314 Ай бұрын

@@megamaz108 It's technically replay editing. A replay file includes in it the seed for queue generation, and an array representing the player's input sequence. I take the replay file of the WR and overwrite the input sequence with the bot's. The bot takes seed -> queue -> generates piece sequence -> translates that into input sequence to put into the replay file. I can technically make changes to the piece sequence or the input sequences if I want, but I usually just let the bot autogenerate the whole thing.

@rolandaugustus1180 Ай бұрын

Brief analysis: rtxile WR run: Jigsaw 1st, TIJS 2nd, Z DPC PCOsaw 1st, TIJZ 2nd, S DPC Jigsaw 1st, TIJZ 2nd, Z SPC PCO 1st, IOSZ 2nd, S DPC Jigsaw 1st, TJLS 2nd, O DPC PCOsaw 1st, IOLS 2nd, S DPC PCOsaw 1st, TILZ 2nd, Z DPC PCO 1st, TOLZ 2nd, Z DPC PCO 1st, JLSZ 2nd, T SPC PCOsaw 1st, TOLS 2nd, J DPC PCO 1st, TOLZ 2nd, Z DPC PCOsaw 1st, TILS 2nd, Z DPC Jigsaw 1st fail to recovery TOJS 2nd, S DPC PCOsaw 1st, TOLZ 2nd, Z DPC Jigsaw 1st, TOLZ 2nd, Z DPC Jigsaw 1st, TIJZ 2nd, O DPC (incomplete) TAS run: Jigsaw 1st, TIJS 2nd, S 3rd, No OJ 4th, IO 5th, No L 6th, ISZ 7th Z replace L 8th, TIJZ 2nd, IOOJL 2L (O 3rd), TSZ 7th IOOJL 2L (O replace T 8th), SZ 5th, No S 6th, IIOJJ 2L (TIJ 7th) TTLSZ dupe 4th, TS 5th, No J 6th, IIOLL 2L (IOL 7th) No IL 4th, LZ 5th, O replace TJ dupe 6th, OLZ 7th J replace I 8th, TIOL 2nd, IOOJL 2L (O 3rd), TSZ 7th Jigsaw 1st, TOLZ 2nd, Z 3rd, IJJJL 2L (IJJLS dupe 4th) S replace J 8th, JLSZ 2nd, Z 3rd, No OL 4th, LZ 5th, O replace IS dupe 6th, TOL 7th IOOJL 2L (O replace T 8th), SZ 5th, No J 6th, OJS 7th IOOJL 2L (O replace T 8th), SZ 5th, No O 6th, IOOJL 2L (TOL 7th) TTLSZ dupe 4th, JS 5th, No Z 6th, TSZ 7th IOOJL 2L (O replace S 8th), TZ 5th, No I 6th, OLS 7th IJSZ 4p 1st, TJLZ 2nd, Z 3rd, No TI 4th, IIOJJ 2L (IJ 5th) TLSZ 2nd, Z 3rd, No IZ 4th, LS 5th, S replace TI dupe 6th

@swng314 Ай бұрын

Wow, thanks for this analysis!

@rolandaugustus1180 Ай бұрын

@@swng314 Np lol Lmk if u need any help w/ PC research stuff Also I wonder what an optimised "humanly conventional" run for this RNG seed would look like (best known 3p/4p setups for 2nd (including qb/oqb), best known DPC/SPC, best known qbs for TSZ 7th). Idk what rules I'd define for 1st under these conditions lol

@swng314 Ай бұрын

@@rolandaugustus1180 I can generate one but it won't be super impressive. The code I have doesn't account for qbs, just basic minimal setup + solution. I'm pretty sure rtxile uses really advanced partial jigsaw knowledge that my code isn't set up to implement. Plus the fact that WR runs by nature have good RNG (e.g. picking a probabilistically "worse" setup but RNG giving a solution queue for a TSD sol regardless)

@rolandaugustus1180 Ай бұрын

@@swng314 Yeah, and I was also thinking that even if you go for the best scoring DPC build on average for each loop, there's still a chance that a "worse" DPC will score higher for that specific bag

@swng314 Ай бұрын

@@rolandaugustus1180 Yup Another example, say you're on 1st and both left and right side jigsaw are buildable. See 7, either choice is the same, but it's possible that for the next bag's queue, one side has a TSD PC sol and one side has missed PC.

@rolandaugustus1180 Ай бұрын

Given how damn optimised this is, rtxile didn't do all that badly on his WR run lmao

@aefad Ай бұрын

That’s a score for 5 pps

@swng314 Ай бұрын

yea I can set it to pretty much any pps so I've just decided to match the blitz WR's pps

@aefad Ай бұрын

@@swng314 Is the new wr faster than the old one?

@swng314 Ай бұрын

@@aefad Yes generally as new WRs come in, they're faster and faster. You can check the WR holder's history: ch.tetr.io/u/rtxile/blitz His current WR is 593 pieces (I'm running 585 pieces for this). His last WR was 570 pieces and I was generating runs for 570 pieces when that was the WR.

@algebruh4185 Ай бұрын

muffled audio head ahh

@swng314 Ай бұрын

oops I guess that's what happens when I have bluetooth headphones in while recording thonk

@smnsr05 Ай бұрын

how does that first j have perfect finesse

@algebruh4185 Ай бұрын

@@smnsr05 ignores softdrop finesse

@smnsr05 Ай бұрын

@@algebruh4185 why did i send this as a reply 💀 thanks for answering though

@algebruh4185 Ай бұрын

is there any way i could be helping with this development or is this just "add features, go wild"

@swng314 Ай бұрын

I may reach out to you about getting some PPB libraries gotta work on some more code before that tho

@13pake Ай бұрын

any desire to do this see 7? I'm guessing that would be quite a bit harder

@swng314 Ай бұрын

I can do see7 DPC with existing jigsaw + 2nd + DPC + 2L+ TSZ 7th libraries. It's not going to score as well as the human WR because a) WR setup choices inherently get better luck results by nature of the fact that it's the WR run. b) My libraries aren't fully fleshed out c) rtxile is just cracked and makes qb decisions, and my libraries are done with a hard split between setup and solution.

@mdqz1_gd Ай бұрын

too close to perfection

@jjeastside Ай бұрын

nah I'd std

@BeCrazy_ Ай бұрын

Skul pure PC spam

@aefad Ай бұрын

That's an efficient bot

@BeCrazy_ Ай бұрын

Actually does it do anything? Still seems like it could just TSD PC whenever before anyway.

@swng314 Ай бұрын

Compare with the last iteration without libraries: kzbin.info/www/bejne/iX-VeaqciLtraKM It finds more tspin PCs (you can line each PC up, they use the same seed), and the score is higher. The tspins the old one found were completely by luck because it wasn't even checking, it was just taking the first PC solution it found.

@BeCrazy_ Ай бұрын

@@swng314 whaeh the so real

@EEDGFE Ай бұрын

may i have the code please...

@BeCrazy_ Ай бұрын

time to learn coding

@swng314 Ай бұрын

I won't be releasing this codebase publicly.

@mongii2 Ай бұрын

aaaaaaaaaaaaaaaaaaaahhhhhhhhhhhhh I is trolling...

@smooshymemes Ай бұрын

i guess rtxile beat tasbot then

@13pake 2 ай бұрын

is this see-11?

@swng314 2 ай бұрын

@@13pake yeah

@ryurai7772 2 ай бұрын

wait but why does the bot soemtimes pauses to think? is it to sync it up with rxtile's gameplay?

@swng314 2 ай бұрын

@@ryurai7772 yes I've set it so it's only allowed to harddrop when he harddrops

@ryurai7772 2 ай бұрын

@@swng314 ooo icic

@crusaderonmeth454 2 ай бұрын

how do i learn freestyle pc like this

@yumm186 2 ай бұрын

flynn from breaking bad ahh narration

@arock9529 2 ай бұрын

IIILJ 2l :wtf:

@iamjorj 2 ай бұрын

pc how

@EEDGFE 2 ай бұрын

Ehats the name of the bot?

@swng314 2 ай бұрын

I don't really have a name for it or even an account, it's just some code slapped together lol

@EEDGFE 2 ай бұрын

@@swng314may I have the code

@EEDGFE 2 ай бұрын

May I have the code then @@swng314

@TST_nb 2 ай бұрын

what's the replay id?

@swng314 2 ай бұрын

There won't be one, I'm definitely not submitting any bot scores to tetrio servers

@918muse7 2 ай бұрын

I've tried this before with CNN without searching. I wonder what is the structure of this NN.

@swng314 2 ай бұрын

It appears this solver still does searching, the NN is a very simple boardstate evaluation used to speed up the search

@918muse7 2 ай бұрын

@@swng314 Oh I see, NN heuristics while searching

@BeCrazy_ 2 ай бұрын

still don't get how you got the bot to have such a human-like movement while its placement are batshit insane

@pepperoni42069-p 2 ай бұрын

people will use anything but minus kelvin pcf

@swng314 2 ай бұрын

part of me dies inside every time I see Rust syntax

@snowman99tetris 2 ай бұрын

What

@swng314 2 ай бұрын

pure PC instead of DPC Neural Net solver instead of using a pregenerated DPC library

@feltheshovel9251 2 ай бұрын

wow cool sequence!

@swng314 2 ай бұрын

@joshinon6980 2 ай бұрын

lets go downstacking

Ең жақсы KZbin

Пікірлер