This is interesting from an optimization perspective too.
The intermittent big jumps are what you need to get out of local optima. You can use gradient ascent locally, but once you stop improving, a big change is needed.
In „The Secret of Our Success“ (aside: everytime I write that title I cringe a little at how self-help-y it sounds), Joseph Henrich conjectures that human divination rituals might have evolved as a way to randomize hunting behavior so as not to become predictable to the animals being hunted for. Not sure how plausible to find it, but I found the thought pretty intriguing.
Not sure if this is from the same book but I've heard something similar about "lunacy" being linked to the full moon.
The reason there is more crime and wild behavior during a full moon is not because of any gravitational effects of the moon etc. Rather, we are descended from tribal peoples who celebrated the full moon with wild parties. Said parties led to intercourse which led to children etc.
Not sure how valid that is but I've always liked the idea.
You can do more shit at night without a fire source when the full moon is out. It can be uncomfortably bright if unaccustomed to perpetual artificial light
I heard we hunted mammoths by moonlight, so mania (which also has a lunar etymology) ie. high energy, adventurousness, aggression are associated with moonlight.
I can't remember the details but apparently they found some engraved mammoth bones which indicated this.
I found this hypotheses in "The Blue Sense" (which gives other interesting related ideas) but it might have an even older origin. Nevertheless, it is an intriguing idea.
Considering the variety of animals that seem to show this behavior, isn't it possible that mammals, specifically humans, are "encoded" with that same random search pattern? I'm not sure how you'd test for it though.
I instantly thought about eye tracking experiments, in which we see lots of small movements around an area, then big jumps across the screen. I guess others have thought about it too.
> Specifically, we use comparisons of maximum-likelihood fit as well as standard deviation analysis and diffusion entropy analysis to show that visual search during language comprehension exhibits Lévy-like rather than Gaussian diffusion.
That is fascinating. Thanks for that.
> However, one could argue, from an evolutionary standpoint, that spe-ciÿc search mechanisms could have been learned and “wired” in order to improve theexploration eciency (e.g., if a salient point is located within a direct vision distance,maximize the probability of straightforwardly moving to that site)
>Ironically, their paper arrived only months after a study co-authored by Viswanathan, da Luz and Raposo showed a catastrophic error in Viswanathan’s original albatross study. The wet/dry sensors they had relied on weren’t dry only when the birds flew: They were also dry when the albatrosses sat on their nests. Some of the longest “flights” in the data set might have actually been protracted visits to the birds’ nests. The study also criticized certain analytical techniques and claimed that albatrosses don’t do Lévy walks after all, and that it was unclear whether many species did.
The answer is immediately in the paragraphs following.
> Sims acknowledges that the conflicting papers “put people on edge.” But his work used more up-to-date mathematical methods to demonstrate that a whole menagerie of marine predators take Lévy walks.
> These improved methods quashed doubts over whether animals ever do Lévy walks. “The amount of data is overwhelming,” said da Luz.
> Sims and Humphries even revisited albatrosses in a subsequent study in 2012. Working with more sophisticated GPS tracking devices, they found that albatrosses did not perform Lévy walks while searching for food above shallow water, where they might have more visual or olfactory information to guide them. But when they flew out over deeper water, the birds did in fact adopt Lévy searches.
So the original work had the right conclusion (Albatrosses do adopt Levy searches) for totally flawed reasons (the data didn't actually support the conclusion, the Levy walks weren't being used for the time periods originally predicted).
Relatedly, a paper was posted to bioRxiv a few days ago describing the results of large scale fruit fly release-and-recapture experiments. They conclude the abstract with "Our field data do not support a Lévy flight model of dispersal, despite the fact that our experimental conditions almost perfectly match the core assumptions of that theory."
Also, Viswanathan's 1996 Nature paper was stupendously wrong due to the key measurements being in error, but instead of retracting it, he and colleagues published a followup - in Nature 2007 - in which they say the albatross flight times are gamma distributed.
18 comments
[ 3.9 ms ] story [ 52.8 ms ] threadThe intermittent big jumps are what you need to get out of local optima. You can use gradient ascent locally, but once you stop improving, a big change is needed.
The reason there is more crime and wild behavior during a full moon is not because of any gravitational effects of the moon etc. Rather, we are descended from tribal peoples who celebrated the full moon with wild parties. Said parties led to intercourse which led to children etc.
Not sure how valid that is but I've always liked the idea.
I can't remember the details but apparently they found some engraved mammoth bones which indicated this.
https://slatestarcodex.com/2019/06/04/book-review-the-secret...
https://journals.aps.org/pre/abstract/10.1103/PhysRevE.79.05...
https://www.academia.edu/1112368/Modelling_gaze_shift_as_a_c...
I'm intrigued.
That is fascinating. Thanks for that.
> However, one could argue, from an evolutionary standpoint, that spe-ciÿc search mechanisms could have been learned and “wired” in order to improve theexploration eciency (e.g., if a salient point is located within a direct vision distance,maximize the probability of straightforwardly moving to that site)
>Ironically, their paper arrived only months after a study co-authored by Viswanathan, da Luz and Raposo showed a catastrophic error in Viswanathan’s original albatross study. The wet/dry sensors they had relied on weren’t dry only when the birds flew: They were also dry when the albatrosses sat on their nests. Some of the longest “flights” in the data set might have actually been protracted visits to the birds’ nests. The study also criticized certain analytical techniques and claimed that albatrosses don’t do Lévy walks after all, and that it was unclear whether many species did.
> Sims acknowledges that the conflicting papers “put people on edge.” But his work used more up-to-date mathematical methods to demonstrate that a whole menagerie of marine predators take Lévy walks.
> These improved methods quashed doubts over whether animals ever do Lévy walks. “The amount of data is overwhelming,” said da Luz.
> Sims and Humphries even revisited albatrosses in a subsequent study in 2012. Working with more sophisticated GPS tracking devices, they found that albatrosses did not perform Lévy walks while searching for food above shallow water, where they might have more visual or olfactory information to guide them. But when they flew out over deeper water, the birds did in fact adopt Lévy searches.
So the original work had the right conclusion (Albatrosses do adopt Levy searches) for totally flawed reasons (the data didn't actually support the conclusion, the Levy walks weren't being used for the time periods originally predicted).
https://www.biorxiv.org/content/10.1101/2020.06.10.145169v1
Also, Viswanathan's 1996 Nature paper was stupendously wrong due to the key measurements being in error, but instead of retracting it, he and colleagues published a followup - in Nature 2007 - in which they say the albatross flight times are gamma distributed.