__ __ _ _____ _ _ _
| \/ | ___| |_ __ _| ___(_) | |_ ___ _ __
| |\/| |/ _ \ __/ _` | |_ | | | __/ _ \ '__|
| | | | __/ || (_| | _| | | | || __/ |
|_| |_|\___|\__\__,_|_| |_|_|\__\___|_|
community weblog
Napkin math
Experimental testing of exploratory mathematics; on Feynman's FOMO/YOLO tradeoff.
"It's hard to study humans because they're often, in effect, solving a more complex version of the problem than the one you've given them. We tell them explicitly this is the distribution, but they still want to figure it out for themselves. Human cognition is often solving something that's more ambiguous and more nuanced than what you're trying to test."
posted by clew on Jun 08, 2026 at 1:55 PM
---------------------------
Nice bit near the end that sounds like humans do piece wise linear approximations of the actual optimal search/stopping algorithms.
posted by clew at 1:56 PM
---------------------------
"What we find in our human participant data is that people are very reluctant to do that even when they get very lucky early on. So even if people are fortunate enough to encounter a really good dish on their first night, they'll keep exploring at least for a little while."
How would one know they got 'lucky'? I'm not sure I follow. Dishes are not rated and ratings are extremely personal.
posted by The_Vegetables at 2:22 PM
---------------------------
The assumption is that each diner has a personal ranking of all the dishes at a particular restaurant. However, they don't know this ranking ahead of time. So the mathematical problem is whether to keep ordering a dish of known quality, or try a new one, which might be better or might be worse.
(The article itself discusses the problems in the assumption, such as the restaurant quality declining, or the diner getting bored.)
posted by zompist at 2:48 PM
---------------------------
"So even if people are fortunate enough to encounter a really good dish on their first night, they'll keep exploring at least for a little while."
Well sure, that first dish raises the likelihood that you've found a really good restaurant and you should try some of their other work. I.e. a restaurant whose probability distribution skews high.
The paper:
The values are assumed to follow a Uniform distribution, with any value between 0 and 1 being equally likely.
oh
posted by away for regrooving at 3:12 PM
---------------------------
But is it a uniform distribution for that restaurant?
posted by Jon_Evil at 4:43 PM
---------------------------
There's a truck, it comes from the laundry on Tuesdays...
but seriously, the uniform distribution is from Feynmans original framing, yesno? From before numerical methods were easy even for Feynman to access. I think that's another case of a computation-thrifty framing that works pretty well.
posted by clew at 5:07 PM
---------------------------
is this within the field of Bistromathics?
In any event, if more research gets funded, I would love to participate.
posted by Artful Codger at 6:55 PM
---------------------------
"So even if people are fortunate enough to encounter a really good dish on their first night, they'll keep exploring at least for a little while."
Well sure, that first dish raises the likelihood that you've found a really good restaurant and you should try some of their other work. I.e. a restaurant whose probability distribution skews high.
Potentially, but really it's intuitive that a person would need to build a dataset for each restaurant before they can judge the quality of their meal. IMO the more interesting question is "how small is that dataset for each restaurant for the average person?" 2 visits? 10? And does that dataset size apply to other similar issues, like parking or online dating, or does each activity get its own dataset size?
Or conversely, it is 1? This could be one explanation why chains are popular for travellers when there are tons of local restaurants that serve similar food and often at better prices and higher quality. They are not willing to risk a novel experience when the dataset is 1.
posted by The_Vegetables at 7:44 AM
---------------------------