Change one variable at a time - good advice? - Page 2

Beginner and pro baristas share tips and tricks.
User avatar
malachi

#11: Post by malachi »

I guess I'm a moron - because I almost never change more than one variable at a time.

To be fair, this is largely because my goal is rarely to get an "acceptable shot" but is instead usually to learn more about the coffee. As such, it works well for me to change a single variable at a time, as I can then taste the results and start seeing patterns in the relationships between these variables and the flavours. In some cases, I am reasonably sure that I'm going to have to change two or more variables (and even how I'm going to have to change them) but changing one at a time and tasting each result educates me better IMHO.


The two exceptions to the "one variable at a time" rule (for me) are:
1 - I'm making drinks for other people
2 - I have a new coffee and I want to quickly get to a "center point" before I start experimenting.
"Taste is the only morality." -- John Ruskin

User avatar
another_jim
Team HB

#12: Post by another_jim »

For the last time, it's never an experiment, it's either fixing a bad shot or exploring the extraction space.

Fixing a Bad Shot: The optimum procedure is the same as searching for the right answer after some initial guess -- you go to your best guess of the correct variables, prorated in some way (Bayesian or Maximum Liklihood) by your degree of certainty in the guess. If you are very certain the shot was much too short, make a big correction in grind, if you also think the shot was way too hot, but not very certain, then make the next shot slightly cooler as well.

Why on earth would you do only one of these at a time? What warrant do you have for suggesting such a silly procedure? There is certainly nothing in statistics or optimization theory that says changes must be made in one dimension at a time. The key is in following the estimated direction of the required change, no matter how many dimensions it involves, but taking careful small steps, then making appropriate corrections.

Exploring the Extraction Space: If a person is a newbie, then they will make a lot of poor shots, and make bad guesses on how to correct them. They will become better if they explore the extraction space. This exploration will involve making shots all over the extraction space, something that will happen in time willy-nilly. But suppose you go about it systematically. Then you would make shots that cover a lattice of points in the extraction space. For instance {20, 26, 32 seconds} * {90C, 92.5C, 95C} * {8 bar, 9 bar, 10 bar} * (13 grams, 16 grams, 19grams} Notice that this is a lattice of 27 shots, and involves doing every combination of variables. You can plot a trajectory that changes one variable at a time in covering this lattice, or you can chart one that changes all four at time. As long as you make all 27 shots, it doesn't matter. In other words, when it comes to exploring the extraction space, the statement "one variable at a time" is meaningless. The key is completely covering the space, whether you do it in 90 degree lines, 60 degree lines, or a drunkard's walk.

This "one variable at time" is one of those pieces of advice that was given ten years ago on alt.coffee, and accepted as conventional wisdom. Nobody has ever given it another thought; even though it's rubbish.
Jim Schulman

User avatar
another_jim
Team HB

#13: Post by another_jim »

dgreen wrote:The only remaining question that still hasn't been answered is whether anyone has successfully pulled 14g shots on a Silvia. It might save me some time to know that.
I once used a pair of Silvias for grinder testing. 14 gram doubles were easier on it like on any other machine. (in the sense of being less likely to channel with less than perfect packing). The Silvia comes stock with a shallow basket, so it's also particularly bitchy at high doses when using that basket.

How well made shots tasted depended on the grinder and coffee. My recall is that with the poorer grinders, the Silvia was particularly kind at low doses.
Jim Schulman

User avatar
timo888

#14: Post by timo888 »

another_jim wrote: ... If someone tries a suit that's three sizes too small, telling them to buy the suit, then go to a tailor and change just one thing at a time, the right sleeve, the left cuff, etc etc, is really, really bad advice. Newbies are prone to do everything wrong on their espresso machine, so telling them to change just one thing at a time is also very bad advice.
...
It's best to start out by specifying a complete set of all the variables that are known to work.
Jim's moron mantrist is a straw man. Of course one wouldn't recommend something as patently absurd as Why don't you try changing the grind and the grind only --it doesn't matter what the water temperature is or how little coffee you're placing in the filter basket. What we morons are recommending is that changes be made to one variable only, with an assumed starting position for all variables that is likely to yield decent results in the cup.

One frequently reads "all else being equal" or "ceteris paribus" in discussions that attempt to isolate and describe the effects of changes made to one variable only. Yet quite a lot of care and focus are required to keep "all else equal" so that changes are made to only one variable. This advice therefore is usually given in the hope that it will encourage the advisee to pay measured attention to the process and not be acting haphazardly. One wants to avoid situations where the advice has been to try, say, a very light tamp, but now, two or three fewer grams of coffee are in the basket too, and the reported results are "it flowed too fast and the taste was thin".

JimG

#15: Post by JimG »

Waaaaaaay back in the day...I wrote quite a bit of technical code for solving multivariable non-linear systems. The method that proved most effective for my problems, IIRC, was one of Newton's contributions. It required making an estimate of the partial derivative of each variable (i.e., hold all but one variable constant and make an estimate of the change to the residual error).

Corrections were made to all variables for each iteration. Recalculation of partial derivatives, however, was also necessary if any of the variables had changed significantly because of the non-linearity of the problem.

This particular method converged very quickly once you were close to the solution, but could be wildly unstable when not close.

Optimizing an espresso shot is analogous. Only after you know the effects of changing each variable independently, i.e. the partial derivatives, are your refinements going to result in a convergent process.

Applied to the subject of this thread, the message is that you should at least determine the "sign" of the partial derivatives: Now that I am close, does an increase in X make the shot better or worse?

Failing to do this introduces the possibility that a good correction to one variable masks a bad correction to another variable, resulting in no change in the shot.

Jim

User avatar
GC7
Supporter ♡

#16: Post by GC7 »

JimG has it about right in my opinion.

Jim - your arguments for espresso along with your suit analogy can only be of use if the variables can be measured and evaluated. Knowing about dosing, temperature and other variables for making espresso and how they can make a difference in the cup is a prerequisite before you can even begin to think about changing multiple parameters akin to your tailor measuring your inseam and arm lengths for your suit alterations. Does a beginner know how to measure and evaluate multiple variables in making espresso? Does the majority of people following this discussion group?

Of course it IS AN EXPERIMENT.

User avatar
malachi

#17: Post by malachi »

Jim... I value your opinion, but you must know that it is just that - an opinion.

A challenge with this site is that it seems to propagate a lot of dogma (believe me, I'm as guilty of it as anyone). The problem with dogma is that it assumes universality.

I'm not going to argue that you're wrong Jim - but rather that both of us are "right."

For me - your approach would result in me drinking a lot of decent or good espresso. But I would learn more slowly. I would still learn, of course, but it would be less efficient for me.

Believe me, I've tried both approaches - and (for me) it's clear that I develop better pattern-recognition skills from the "moron" approach. Not only do I develop better understanding of the impact of changes to variables - I also start to see the relationships and inter-dependencies between these variables.

Obviously, it might not work for other folks. And it almost certainly won't work if your goals are different than mine.
"Taste is the only morality." -- John Ruskin

User avatar
another_jim
Team HB

#18: Post by another_jim »

What's wrong with this shot? How do I fix it? That is the context of this discussion.

JimG illustration of Newton's steepest descent method* is a good one.

An experienced person will somewhat know what's gone wrong, i.e. will have some idea of the partial derivatives of all the variables. His best strategy is to make a change in all the variables' estimated directions.

An inexperienced person will have no clue about what's gone wrong, and is stuck with a brute force search of the extraction space. That can be done one, two three or X variables at a time, as long as the entire space is searched.

In either case, "one variable at a time" makes no sense.

So, whenever I think a shot is, for instance, a little too slow and a little too low dosed, I'm going to continue making both adjustments at the same time. The rest of you are welcome to flip a coin and waste a few shots getting there "one variable at a time." :P


*BTW: What's changed in the last 300 years, mostly courtesy of Gauss, is that one can generate a least square error like estimate of the partial derivatives, based on the data being used, and assign a 0 to 100% degree of confidence to each estimate. Then one prorates how far one corrects in each direction by using the product of the derivative and its degree of confidence. This makes the search far from the target a lot more stable and faster than the pure Newtonian search.
Jim Schulman

zin1953

#19: Post by zin1953 »

another_jim wrote:So, whenever I think a shot is, for instance, a little too slow and a little too low dosed, I'm going to continue making both adjustments at the same time.
As well you should, Jim. But what works for you may not work for someone else.

You presume that the individual pulling the shot knows that the shot is BOTH "a little too slow" AND "a little too low dosed." And in that case, it certainly would make sense for you (in this case) to alter both variables at once. And I have no doubt whatsoever that, due to your experience, you could indeed tell that the shot was both "a little too slow and a little too low dosed." You might even be able to tell that the temperature was off a bit too, and correct for all three factors on your second pull.

Let us not lose sight of the fact this thread was split off from an original post by someone with far less experience than you. As such, the individual -- any newbie individual -- may not know that the shot was too slow and the dose was low. All he or she knows, from a starting point of 14g/25 sec/50ml volume is that "This isn't right." It may not be total crap (then again, it may), but it's clearly -- to the palate of the newbie -- not right. What should he/she do? Go straight to an 18g/38 sec/35ml shot?

I don't think so.

There is always more than one way to skin the proverbial cat, more than one way to make an espresso. We all have our own palates, our own individual tastes, and the "right" way to pull a shot varies -- not only among individuals, but with a single individual (how many adjustments need to be made throughout a day, or several days, as the weather changes, as the beans age, etc., etc.?). What works for one individual may not work for another, and what an "expert" can tell from tasting an "off" shot is going to be different from what someone with far less experience could tell from tasting that same shot.

There are many different ways to get to there. Each person finds their own path -- it may turn out that it's (all or in part) the same path others have taken, but there is no substitute for making the journey yourself, for learning the variables and what role(s) they play, so that -- in the future -- you will be able to more quickly recognize and diagnose the problem(s), and make multiple adjustments "on the fly" as it were.

Cheers,
Jason
A morning without coffee is sleep. -- Anon.

User avatar
kschendel

#20: Post by kschendel »

another_jim wrote: So, whenever I think a shot is, for instance, a little too slow and a little too low dosed...
And what is one to do if one doesn't know that it's too low dosed? just that it doesn't taste good?

You made noises about "exploring the extraction space", which is fine. Now, why does one do that? to find out what effects (if any) the different variables have. You can do that by either a) changing one thing and holding the others constant, or b) doing a whole slew of experiments that cover the space, and then mapping the results, looking to learn what variables do stuff and which ones don't. (a) and (b) are exactly equivalent. You do the same total number of experiments either way. Most engineers and researchers prefer (a) because you get partial knowledge before the full series ends.

If you already know which variables do what and which ones don't, of course changing one thing at a time is silly. That is not the context in which I usually see "change one thing at a time" advised on these forums.

If you go into any reputable research or performance lab (at least the ones I've been associated with over the years), and say "changing one thing at a time (during exploratory research) is for morons", I guarantee you'll have one personal variable changed: your height off the ground. From the window.