The method debate thread

[G2013]

However, recently I was re-generating certain 'bad cases' of LMCF algorithms, and cube explorer spit out something that looked awful, but ended up pretty good, I call it the sexE move

Expanded Cell Ver. of Max's Algs

 

[Tao Yu]

Well heck. an you go more into your opinion on method development
?

I feel like it's just the same story over and over again. Before I open up any method description I know it will just be solving things in some random order. And methods just never go anywhere, it always just ends up with the creator sticking with it and having like a 14 second average. I just think there are more interesting things in cubing. I have nothing against other people being interested it though.

Indeed LR does look like a bad step. However, recently I was re-generating certain 'bad cases' of LMCF algorithms, and cube explorer spit out something that looked awful, but ended up pretty good, I call it the sexE move:
U' L E' L' U' L E L'
This is basically the sexy move except some of the U moves are replaced with E moves. I thought it would be crap. But on a properly tuned cube, after some drilling, I can consistently do it in 0.90 seconds (8.88tps) and I'm a very slow turner (3-4 tps average on full solves). I think an expert could do this in 0.50 seconds. The same mirror exists on the R/U/E set. This weird trigger opens up a large number of possibilities for move reduction in LMCF and also (possibly) makes certain methods like Mehta that require E moves somewhat more practical. I'm curious if a fast turner here can try this trigger, try timing 10 repetitions of the sexE move and then divide by 10 to see how long it takes you. Personally I find the ULEL version easier than the URER version.

As someone developing a corners first method, you really need to have a good look into the commutators used in 3-style. BLD solvers have been using algs like this for years. Your analysis of the speed is correct, in fact it's probably even sub 0.4 consistently.

I think the problem with many of these algs for sighted speedsolving is that they don't move enough pieces. It's pretty difficult to use 3-cycles like the one you gave to move more than one or two pieces in a way that you would want. Moving one or two pieces at a time just isn't good enough for sighted solving.

Edit: For your own sake, I wouldn't recommend calling it the sexE move. BLD solvers will never stop making fun of you since it's one of the most common algs in BLD solving.

Edit2: To clarify my point about not moving enough pieces at a time. F2L also solves two pieces at a time, but I think it's different for the following reason:

You can solve two edges (FR and FL) in one alg using L' U L E' L' U' L E. However, this only works if one of the edges you need to solve is in the middle layer at the very start. If it's not there already, you have to put it there first.

F2L, on the hand, solves two pieces, while allowing them to start from any position at the very start.

 

[Athefre]

I feel like it's just the same story over and over again. Before I open up any method description I know it will just be solving things in some random order. And methods just never go anywhere, it always just ends up with the creator sticking with it and having like a 14 second average. I just think there are more interesting things in cubing. I have nothing against other people being interested it though.

With a few exceptions of course.

 

[PapaSmurf]

With method development, I agree mostly with Tao's take. If your aim is a meme, make stupid methods. If your aim is an equal-or-better-than-big-3 method, you're gonna have to come up with something better than "let's do cross but using any mismatched corners and call it BeltFOP" or similar. Mehta is a method which I think has at least some potential (and is almost certainly the best method to be developed in the past year), but even that has gone through a lot of refinement and definitely isn't better than the big 3, only equal at best. It's frustrating, but most good method ideas have gone already and the mediocre ones aren't far behind. Even then, unless it's for method science, mediocre ones aren't really worth it. Read the "New Method/Substep/Concept" thread from start to finish (if you can be bothered). What you'll find is a lot of repeats of the same (bad) ideas and a few almost good ideas that were dropped because they weren't as good as the big three and therefore their value is lost instantly. It shows you how fruitless the search for better methods has been. It's not to no avail though. More significant ideas include WaterRoux (an almost good method that could be pushed a bit further) and LMCF (an almost good method that could be pushed a bit further) and Zipper (an actually good method that not enough people know about). The only method that I can think of that has been criminally underdeveloped is 42. If people learnt the algs and the recog, I reckon there would be a potential minor increase in Roux speed. Taking these ideas could help slightly change the path of speedsolving, but not create a whole new meta.

Taking all this into account, the method that averages 35 moves and 10 TPS is, unfortunately, probably not going to come. Techniques such as L3C reduction for ZZ (which will potentially come with an automised cycle union programme), 42 for Roux and people noticing Zipper for CFOP are where the next breaks lie, and they most likely won't be large enough to give anyone who uses them a significant (>0.5) competitive advantage.

The only bit of hope I see is if someone could break down what makes a method good to the most fundamental level, then optimise using a computer programme to get a truly objectively optimal method. Even then, it's a two-edged sword, as once the most optimal method is found, the whole idea of method development dies, although it would be a fun project for anyone who wants to attempt it. Even if it didn't find the most optimal method, rather a more optimal method, it would put the barrier of making new methods so high as you'd need to understand machine learning, speedsolving in depth, etc.

TL;DR method development is unfortunately probably not going to progress much past minor optimisations, which is a shame. The best chance of a better method is machine learning/AI. Even then, that would either kill the whole of method development or make the barrier of entry so high it would be pointless anyway.

If you couldn't tell, I've thought about it a lot. There are only so many ways to solve 20 pieces and to find novel and (more) speed and move efficient ways is very difficult.

 

[Nir1213]

With method development, I agree mostly with Tao's take. If your aim is a meme, make stupid methods. If your aim is an equal-or-better-than-big-3 method, you're gonna have to come up with something better than "let's do cross but using any mismatched corners and call it BeltFOP" or similar. Mehta is a method which I think has at least some potential (and is almost certainly the best method to be developed in the past year), but even that has gone through a lot of refinement and definitely isn't better than the big 3, only equal at best. It's frustrating, but most good method ideas have gone already and the mediocre ones aren't far behind. Even then, unless it's for method science, mediocre ones aren't really worth it. Read the "New Method/Substep/Concept" thread from start to finish (if you can be bothered). What you'll find is a lot of repeats of the same (bad) ideas and a few almost good ideas that were dropped because they weren't as good as the big three and therefore their value is lost instantly. It shows you how fruitless the search for better methods has been. It's not to no avail though. More significant ideas include WaterRoux (an almost good method that could be pushed a bit further) and LMCF (an almost good method that could be pushed a bit further) and Zipper (an actually good method that not enough people know about). The only method that I can think of that has been criminally underdeveloped is 42. If people learnt the algs and the recog, I reckon there would be a potential minor increase in Roux speed. Taking these ideas could help slightly change the path of speedsolving, but not create a whole new meta.

Taking all this into account, the method that averages 35 moves and 10 TPS is, unfortunately, probably not going to come. Techniques such as L3C reduction for ZZ (which will potentially come with an automised cycle union programme), 42 for Roux and people noticing Zipper for CFOP are where the next breaks lie, and they most likely won't be large enough to give anyone who uses them a significant (>0.5) competitive advantage.

The only bit of hope I see is if someone could break down what makes a method good to the most fundamental level, then optimise using a computer programme to get a truly objectively optimal method. Even then, it's a two-edged sword, as once the most optimal method is found, the whole idea of method development dies, although it would be a fun project for anyone who wants to attempt it. Even if it didn't find the most optimal method, rather a more optimal method, it would put the barrier of making new methods so high as you'd need to understand machine learning, speedsolving in depth, etc.

TL;DR method development is unfortunately probably not going to progress much past minor optimisations, which is a shame. The best chance of a better method is machine learning/AI. Even then, that would either kill the whole of method development or make the barrier of entry so high it would be pointless anyway.

If you couldn't tell, I've thought about it a lot. There are only so many ways to solve 20 pieces and to find novel and (more) speed and move efficient ways is very difficult.

Maybe if we cant really think of good methods that can surpass the big 3 anymore, what about improving one of the big 3 methods? If that doesnt work, then i dont know really.

 

[Tao Yu]

I still haven't given up on the idea that there might be stuff out there that we haven't discovered yet, but I just don't find most approaches these days interesting enough to follow in any detail. Even stuff that is potentially good just ends up nowhere.

I'm with @PapaSmurf that I think software is the way ahead. I'd be a lot more interested in method development if we had some of the following programs:

• A machine learning model for choosing the best alg for a case. In particular, I would like to know if we could use a ton of smart cube data to figure out what sequences of moves on a cube can be performed quickly or get an idea of the ideal way to transform algs to make them as fast as possible.
• A general algset generator, which can generate algs from one step to another automatically. HARCs may be good enough for this already, but I'm not too familiar with it. In any case, things can probably always be improved, and combining this with better models for determining whether an algorithm is good would be interesting.
• A smart cube program which can help you test methods without knowing all the algs. e.g. if you were testing ZB, the smart cube would keep track of what step you are on so that when you get to last slot, the program would show you the ZBLS alg, and when you get to ZBLL, it would automatically show you the ZBLL alg

 

[PapaSmurf]

I still haven't given up on the idea that there might be stuff out there that we haven't discovered yet, but I just don't find most approaches these days interesting enough to follow in any detail. Even stuff that is potentially good just ends up nowhere.

I'm with @PapaSmurf that I think software is the way ahead. I'd be a lot more interested in method development if we had some of the following programs:

• A machine learning model for choosing the best alg for a case. In particular, I would like to know if we could use a ton of smart cube data to figure out what sequences of moves on a cube can be performed quickly or get an idea of the ideal way to transform algs to make them as fast as possible.
• A general algset generator, which can generate algs from one step to another automatically. HARCs may be good enough for this already, but I'm not too familiar with it. In any case, things can probably always be improved, and combining this with better models for determining whether an algorithm is good would be interesting.
• A smart cube program which can help you test methods without knowing all the algs. e.g. if you were testing ZB, the smart cube would keep track of what step you are on so that when you get to last slot, the program would show you the ZBLS alg, and when you get to ZBLL, it would automatically show you the ZBLL alg

Smart cubes do give a lot of potential for data collection. Can someone do a PhD on speedsolving and smart cubes etc.? Please and thanks.

The best way to find the best alg is to somehow model human hands perfectly with some sort of biomechanics simulator engine, then run tons of algs through it using already existing alg genorators. There will probably be some funky algs, but they would be good. Data from smartcubes would be helpful to show what kinda moves are worse or better, although that could lead to stagnant finger tricks (if we did this 5 years ago there would be even fewer LUE comms or algs with Bs).

I think HARCS can do it, but I'm the same with my knowledge.

That's a good idea which I don't think would be that difficult to implement. Would also make learning algs in context easier.

I know my last post was a bit defeatist, but I am hopeful that a better method is out there, just not very optimistic about it.

 

[dudefaceguy]

Before I open up any method description I know it will just be solving things in some random order.

I always wanted to make a random method generator which would take basic steps, jumble them up, and give you a method to try on your next solve just for a fun challenge. You'd get something like:

Orient corners
Solve a 2x2x3 block
Solve columns
Orient edges
Solve remaining F2L edges
PLL

Anyhow, I really like trying different methods because I'm not focused on speed. New steps can be a lot of fun to mess around with, especially if you just read the steps and figure out how to solve them yourself instead of looking up the algs. SSC is one of my favorites for this. And of course Heise is tons of fun. I just love that there are so many ways to think about solving.

But of course, if your goal is speed then none of that matters and you're not really going to beat the top methods. I don't even think machine learning and robots would get us very far, since human brain recognition is the real bottleneck. We already have thousands of humans running aggressive optimization routines on themselves.

 

[Silky]

With method development, I agree mostly with Tao's take. If your aim is a meme, make stupid methods. If your aim is an equal-or-better-than-big-3 method, you're gonna have to come up with something better than "let's do cross but using any mismatched corners and call it BeltFOP" or similar. Mehta is a method which I think has at least some potential (and is almost certainly the best method to be developed in the past year), but even that has gone through a lot of refinement and definitely isn't better than the big 3, only equal at best. It's frustrating, but most good method ideas have gone already and the mediocre ones aren't far behind. Even then, unless it's for method science, mediocre ones aren't really worth it. Read the "New Method/Substep/Concept" thread from start to finish (if you can be bothered). What you'll find is a lot of repeats of the same (bad) ideas and a few almost good ideas that were dropped because they weren't as good as the big three and therefore their value is lost instantly. It shows you how fruitless the search for better methods has been. It's not to no avail though. More significant ideas include WaterRoux (an almost good method that could be pushed a bit further) and LMCF (an almost good method that could be pushed a bit further) and Zipper (an actually good method that not enough people know about). The only method that I can think of that has been criminally underdeveloped is 42. If people learnt the algs and the recog, I reckon there would be a potential minor increase in Roux speed. Taking these ideas could help slightly change the path of speedsolving, but not create a whole new meta.

Taking all this into account, the method that averages 35 moves and 10 TPS is, unfortunately, probably not going to come. Techniques such as L3C reduction for ZZ (which will potentially come with an automised cycle union programme), 42 for Roux and people noticing Zipper for CFOP are where the next breaks lie, and they most likely won't be large enough to give anyone who uses them a significant (>0.5) competitive advantage.

The only bit of hope I see is if someone could break down what makes a method good to the most fundamental level, then optimise using a computer programme to get a truly objectively optimal method. Even then, it's a two-edged sword, as once the most optimal method is found, the whole idea of method development dies, although it would be a fun project for anyone who wants to attempt it. Even if it didn't find the most optimal method, rather a more optimal method, it would put the barrier of making new methods so high as you'd need to understand machine learning, speedsolving in depth, etc.

TL;DR method development is unfortunately probably not going to progress much past minor optimisations, which is a shame. The best chance of a better method is machine learning/AI. Even then, that would either kill the whole of method development or make the barrier of entry so high it would be pointless anyway.

If you couldn't tell, I've thought about it a lot. There are only so many ways to solve 20 pieces and to find novel and (more) speed and move efficient ways is very difficult.

Any specific thoughts on methods that are worth developing/ methods which you'd consider to be as good as the Big4 ?

 

[PapaSmurf]

Mehta, LEOR, Zipper and 42 are the ones that come to mind and I don't think any others are that great or bring anything new.

 

[Silky]

Mehta, LEOR, Zipper and 42 are the ones that come to mind and I don't think any others are that great or bring anything new.

I think that LMCF has huge potential. It seems to be a pretty underrated/overlooked method. One significant advantage that LMCF has is it's flexibility. The fact that you can adjust the method to a scramble relatively easily seems like it could be very useful ( switching between LMCF, Waterman, Roux, WaterRoux(?) ).

 

[PapaSmurf]

LMCF is a method that will occasionally give crazy singles, otherwise meh. Roux is most likely as optimised as corners first will get, although if someone could fix WaterRoux L7E, then that could be equally good from what I know, although I don't know a lot about it.

 

[carcass]

I just finished learning Oll and Pll, I know the algs pretty well, too. I was looking at other sets, but which is the best to learn? WV, BLE, or COLL minus sune and antisune?

 

[PapaSmurf]

This isn't the place for it, but just get OLL and PLL down. Don't distract yourself from the rest of the solve with LL for now.

 

[Deleted member 55877]

I just finished learning Oll and Pll, I know the algs pretty well, too. I was looking at other sets, but which is the best to learn? WV, BLE, or COLL minus sune and antisune?

Optimize OLL and PLL recognition, algs, and fingertricks

 

[Klaudiusz Szyprocinski]

ZB method (short for Zborowski-Bruchem after its proponents, Zbigniew Zborowski and Ron van Bruchem) is a very efficient but algorithm-intensive method.

There are three steps in ZB:
- F2L-1 (F2L without last pair)
- ZBLS (Zborowski-Bruchem Last Slot)
- ZBLL (Zborowski-Bruchem Last Layer)

There are 306 ZBLS algorithms (for inserting the last slot while orienting last layer edges) and 472 ZBLL algorithms (for orienting the last layer corners and permuting whole LL).

ZB is 45.43 moves on average. Roux is 48 and CFOP is 55. That is a huge difference.
You are not restricted to doing cross and then F2L pairs. You can solve F2L-1 however you like, including blockbuilding, pseudoslotting, keyhole etc.
You can minimize the amount of ZBLS needed by learning efficient edge-control. You can also expand by learning VH, RLS, VLS and many more.
Once you know full ZBLL you are already an animal in learning new algs. You can now continue your path and learn new exciting stuff.

This method was originally proposed in 2002. Twenty years have passed and new generation of speedcubers with people like Tymon are constantly proving that ZBLL is essential at the higher level. He can one-look in inspection whole F2L-1 stage and solve the rest of the cube in one or two more looks in a fast and efficient manner.

 

[Eli Apperson]

ZB method (short for Zborowski-Bruchem after its proponents, Zbigniew Zborowski and Ron van Bruchem) is a very efficient but algorithm-intensive method.

There are three steps in ZB:
- F2L-1 (F2L without last pair)
- ZBLS (Zborowski-Bruchem Last Slot)
- ZBLL (Zborowski-Bruchem Last Layer)

There are 306 ZBLS algorithms (for inserting the last slot while orienting last layer edges) and 472 ZBLL algorithms (for orienting the last layer corners and permuting whole LL).

ZB is 45.43 moves on average. Roux is 48 and CFOP is 55. That is a huge difference.
You are not restricted to doing cross and then F2L pairs. You can solve F2L-1 however you like, including blockbuilding, pseudoslotting, keyhole etc.
You can minimize the amount of ZBLS needed by learning efficient edge-control. You can also expand by learning VH, RLS, VLS and many more.
Once you know full ZBLL you are already an animal in learning new algs. You can now continue your path and learn new exciting stuff.

This method was originally proposed in 2002. Twenty years have passed and new generation of speedcubers with people like Tymon are constantly proving that ZBLL is essential at the higher level. He can one-look in inspection whole F2L-1 stage and solve the rest of the cube in one or two more looks in a fast and efficient manner.

For a method to be "better", there are a lot more factors then just speed. "Faster" might be a better word, though I don't know that that is even true. ZB is very algorithm intensive, making it harder to learn the vanilla cfop oll+pll, and this worse in the aspect of how quickly you can learn and progress. Second, recognition time for zb is quite a bit worse then cfop, and takes a lot more practice. I don't think there is a significant difference in OLL+PLL average times compared to ZBLS+ZBLL. I could be wrong, if I am someone please correct me.

 

[Mattecuber]

I have a question, wouldn't be easier to do ONLY VHLS on last slot? It's only 15 algs when you have a pair ready to be inserted and 15 algs when you have a three mover.

 

[zzcuberman]

For a method to be "better", there are a lot more factors then just speed. "Faster" might be a better word, though I don't know that that is even true. ZB is very algorithm intensive, making it harder to learn the vanilla cfop oll+pll, and this worse in the aspect of how quickly you can learn and progress. Second, recognition time for zb is quite a bit worse then cfop, and takes a lot more practice. I don't think there is a significant difference in OLL+PLL average times compared to ZBLS+ZBLL. I could be wrong, if I am someone please correct me.

With the zbls you can predict what case it'll be on 3rd pair pretty easy

I have a question, wouldn't be easier to do ONLY VHLS on last slot? It's only 15 algs when you have a pair ready to be inserted and 15 algs when you have a three move

I have a question, wouldn't be easier to do ONLY VHLS on last slot? It's only 15 algs when you have a pair ready to be inserted and 15 algs when you have a three mover.

Wastes way to many moves. Zbls is far superior

 

[Thom S.]

Funny, you said the exact opposite last week.