Quantifying Lookahead

[obelisk477]

As I was drilling LL today, I decided to do so without inspection to help improve OLL recognition. This got me thinking that there might be a good way to quantify how good at lookahead you actually are with a few time trials. A session of no inspection cross drilling, no inspection F2L (only one pair!), no inspection OLL and no inspection PLL all added together would give you a decent idea of how slow you would be if you were unable to lookahead at any point during a solve. And another session with inspection for all of the above added together again would basically just be an average really only considering TPS. Then you take your approximate global average and see where you fall in terms of percentile within the range between the two calculated averages.

Note: The reason you would do one f2l pair is mainly pragmatic - you can just multiply by 4 to get close enough.

Example:

Without Inspection
Cross: ~4
F2L: (~2.75 * 4) = ~11
OLL: ~3.3
PLL: ~3.3
Total: 21.6

With Inspection:
Cross: ~2.75
F2L: (~2.2 * 4) = ~8.8
OLL: ~2.7
PLL: ~2.7
Total: 16.95

Global Average: 19.1

Percentile: 1 - ((19.1-16.95)/(21.6-16.95)) = 53.76%

I suppose you could either view it as a percentile, or how close you are to just pure TPS. But the fun thing is it can be compared across the spectrum of cubers. You can 'compare' your lookahead to cubers both faster and slower than you using this metric.

p.s. I just hacked this together to avoid writing a paper. I realize theres a flaw or two here or there, but I would love some thoughts, feedback, or even examples of what your percentile is! Also, maybe some suggestions as well for the name of it other than 'percentile'. The name is roughly accurate, but it would be cooler to call it 'lookahead effectiveness' or something.

 

[GuRoux]

why not just the (with inspection total minus the without inspection total)/global average. this would give you the amount of time you spend recognizing for every second you solve.

 

[obelisk477]

why not just the (with inspection total minus the without inspection total)/global average. this would give you the amount of time you spend recognizing for every second you solve.

No, your number here would give you the maximum amount of time you should ever spend on recognition per second you solve, because the 'without inspection total' is calculated by considering the worst case scenario - namely, that you solve each step of the cube completely blind. To get the figure you mentioned, it would be (global minus w/ inspection) divided by global. Which can also be a useful figure.

And of course, in your case this could translate over to Roux with FB, SB, etc.

 

[GuRoux]

No, your number here would give you the maximum amount of time you should ever spend on recognition per second you solve, because the 'without inspection total' is calculated by considering the worst case scenario - namely, that you solve each step of the cube completely blind. To get the figure you mentioned, it would be (global minus w/ inspection) divided by global. Which can also be a useful figure.

And of course, in your case this could translate over to Roux with FB, SB, etc.

oh yes, you're right.

 

[pdilla]

Huh, I've done time trials with and without inspection to see what kind of deviations I would get, but never thought of really quantifying the individual gain/loss per step like this. This would be a very interesting experiment for the community to attempt.

 

[theROUXbiksCube]

I also need to try this, maybe it'll improve my times because my lookahead sucks.
But don't you just reach a point sometimes where you can't stop looking ahead because your tps is "perfectly" in sync with your look ahead and its all fluid?

 

[Rocky0701]

But don't you just reach a point sometimes where you can't stop looking ahead because your tps is "perfectly" in sync with your look ahead and its all fluid?

Yes you do, however by trying these exercises, I believe that it will improve look ahead because it forces your mind to be able to search for pieces quickly and recognize cases with ease.

 

[obelisk477]

I also need to try this, maybe it'll improve my times because my lookahead sucks.
But don't you just reach a point sometimes where you can't stop looking ahead because your tps is "perfectly" in sync with your look ahead and its all fluid?

Yes but I've found that in those times where your lookahead is perfectly in sync with your TPS, your TPS is ever so slightly slower than just really quick blind alg execution, because you've found the right point in your TPS where your lookahead can keep up. That's why 'slow solving' helps sometimes, I think, because it can help you find that balance point. So if that's true, you can never perfectly reach that 'with inspection' total number unless your alg-TPS just improves by itself, in which case you would need to re-calculate.

 

[pdilla]

Yes but I've found that in those times where your lookahead is perfectly in sync with your TPS, your TPS is ever so slightly slower than just really quick blind alg execution, because you've found the right point in your TPS where your lookahead can keep up.

Unless you're faz

 

[Petro Leum]

this isn't method neutral, so you can't quite compare.

 

[guysensei1]

this isn't method neutral, so you can't quite compare.

You can compare with people who use the method you use.

 

[Petro Leum]

You can compare with people who use the method you use.

true, but that's not optimal. after all, lookahead isnt exclusive to a method, but rather just a skill in cubing

 

[obelisk477]

this isn't method neutral, so you can't quite compare.

I'm not quite sure how this isn't method neutral. For Roux you could just do the same thing with the substeps, and subdivide based on your skill level. So it would be FB, SB, CMLL, EO, UL/UR, and (L3E? idk I dont use Roux). But if you can't plan SB fully for example, you would time SB-1 and SB separately. And so on. This isn't too different from CFOP; people who use 4LLL would have to subdivide the last two steps into four.

 

[sneaklyfox]

Interesting. I did the experiment using Ao12 for each substep.

Spoiler: Calculation

Without Inspection
Cross: 4.97
F2L: (2.8 * 4) = 11.2
OLL: 2.45
PLL: 2.68
Total: 21.3

With Inspection:
Cross: 1.57
F2L: (1.46 * 4) = 5.84
OLL: 1.84
PLL: 1.86
Total: 11.11

Global Average: 15.4

Percentile: 1 - ((15.4-11.11)/(21.3-11.11)) = 58%

My percentile was 58%. However, the real number is probably worse because this doesn't take into account the extra time it takes to start and stop the timer between each substep.

 

[obelisk477]

Nice! Its kinda hard to tell how that compares, I guess, without a bunch of other people doing it as well. I'm not sure you could approximate a theoretical average percentile for most cubers - its just gotta be brute force. My guess (and its only a guess) is that most people sub-25 would fall in the 40%-70% range, with it being difficult to get much over 75%.

 

[tseitsei]

Why would you do cross without inspection too?

That doesn't require lookahead since you can plan it during 15s inspection...

Other than that it's an interesting idea

 

[confusedcuber]

From average of 5 of each.

Spoiler: Calculation

Without Inspection
Cross: 6.35
F2L: (3.78 * 4) = 15.12
OLL: 3.33
PLL: 4.29
Total: 29.09

With Inspection:
Cross: 4.24
F2L: (1.75 * 4) = 7
OLL: 1.92
PLL: 1.94
Total: 15.1

Global Average: 28.5

Percentile: 1 - ((28.5-15.1)/(29.09-15.1)) = 4.2%

4.2%.... Learnt 2 things: 1: I don't look ahead much at all, and it would help if I did. 2. Really need to work on cross planning and PLL recognition.

 

[TDM]

How can you do the 'without inspection' part? I always influence or look ahead to the next step in some way, so I found it impossible to not inspect anything...

 

[confusedcuber]

How can you do the 'without inspection' part? I always influence or look ahead to the next step in some way, so I found it impossible to not inspect anything...

Well its easy for cross. For F2L, you could use a cross solution finder, solve the cross not looking at cube, 1 pair, scramble and repeat.

For last layer i just spammed LL algs without looking untill i didnt know what it was anymore.

Was easier for me cus im not as good funnily enough.

 

[sneaklyfox]

How can you do the 'without inspection' part? I always influence or look ahead to the next step in some way, so I found it impossible to not inspect anything...

qqtimer has last layer scrambles under 3x3 substeps. For F2L and PLL, I solved the step previous without looking at the cube. For F2L, sometimes I might unintentionally notice another corner or something though I try not to pay any attention to it so sometimes I also do a cube rotation before the next pair.