5 ms·

Could you please elaborate? Were you referring to the stories told in his book, or were you actually one of Feynman's students? Here's the written story of Fey

by Nevaeh 14y ago

Could you please elaborate? Were you referring to the stories told in his book, or were you actually one of Feynman's students? Here's the written story of Feynman's experimentation with ants if anyone is interested: http://books.google.com/books?id=7papZR4oVssC&pg=PA94&lpg=PA94&dq=feynman+ants&source=bl&ots=esR_gekP_0&sig=f1vmF-eG6OUR78uEwvHc_rHvNqs&hl=en#v=onepage&q=feynman%20ants&f=false http://books.google.com/books?id=7papZR4oVssC&pg=PA94&#3...

ChuckMcM 14y ago
Those are the ones. He came to USC and as a guest lecturer for my quantum mechanics class and recounted this story. I really like it because it crystallized in me the understanding that this is how you figure out systems for which there is no manual or documentation. Observe the working system, inject change in various ways, measure the effects of that change. Rinse and repeat and learn. Once at NetApp there was a question about whether or not we were wasting all of our time randomly seeking the drives (drive seeks being of course the worst if you're looking to improve performance). But the code paths were very complex, asynchronous, and difficult to model. We figured it out by injecting random seeks which was pretty easy to do. You could flood the disk driver with random seeks and get the 'floor' of performance, then you could back off on the random seek injection to see where performance started to pick up. That told us that counter to our thinking we weren't actually being plagued by seeks, the performance issue was elsewhere in the system. That the ants had, through natural selection, derived a congestion control system that is very close to the TCP one is a fun result because it suggests that the TCP one is probably close to optimal. (Ants have been working on it a lot longer)
- corin_ 14y ago
  Does a long period of evolution actually imply it is "probably close to optimal"? I don't know a huge amount about evolution, but my two lines of logic would be: First, given the huge diversity of life on Earth, how can we argue that any species is "optimal". Secondly, the same argument could be made of life 500 million years ago, that it had been evolving for so long that it must be optimal, no?
  - wzdd 14y ago
    It's actually a common misunderstanding of evolution to assume that it is optimising anything other than suitability to a particular environment (see http://en.wikipedia.org/wiki/Objections_to_evolution#Defining_evolution http://en.wikipedia.org/wiki/Objections_to_evolution#Definin...). So no, it doesn't make sense to use evolution as evidence that suitability in one domain implies suitability in another domain, especially when the domains have almost nothing in common except a shared requirement for management of scarce resources in a very abstract sense.
    - hnriot 14y ago
      I don't think you are seeing the parallels here. When you look closer you'll see that the domains are different but the problems are very similar and not so abstract, which explains the high degree of simularity and effectiveness of the congestion control algorithm. I would also take issue with what you see as a common misunderstanding. Evolution is, by it's very definition a process which adapts to a specific environment. Nothing more, no design, but optimal for any given environmental condition. When the environment changes rapidly, the pace of evolution increases, and vice versa. I don't think this is commonly misunderstood at all. Some people will always have no clue and of course there are those that think the fossils were planted by god, but aside from the out of band idiots, evolution is commonly correctly understood.
  - rcthompson 14y ago
    For any living system that has had sufficient time to adapt to its environment, it can generally be assumed that the system has reached (or nearly reached) a local optimum for that environment, for the same reason that if you let a marble roll around on a surface for a long enough time, you can generally assume that it will come to rest at a local minimum. Of course, just as the local minimum of the marble is defined by the surface, the local optimum in the evolution of a trait is defined relative to the current environment. If the environment is rapidly changing, then you cannot assume near-optimality, just like you cannot assume the marble will come to rest on a constantly undulating surface.
    - 14y ago

by Nevaeh 14y ago

ChuckMcM 14y ago

Those are the ones. He came to USC and as a guest lecturer for my quantum mechanics class and recounted this story. I really like it because it crystallized in me the understanding that this is how you figure out systems for which there is no manual or documentation. Observe the working system, inject change in various ways, measure the effects of that change. Rinse and repeat and learn. Once at NetApp there was a question about whether or not we were wasting all of our time randomly seeking the drives (drive seeks being of course the worst if you're looking to improve performance). But the code paths were very complex, asynchronous, and difficult to model. We figured it out by injecting random seeks which was pretty easy to do. You could flood the disk driver with random seeks and get the 'floor' of performance, then you could back off on the random seek injection to see where performance started to pick up. That told us that counter to our thinking we weren't actually being plagued by seeks, the performance issue was elsewhere in the system. That the ants had, through natural selection, derived a congestion control system that is very close to the TCP one is a fun result because it suggests that the TCP one is probably close to optimal. (Ants have been working on it a lot longer)

corin_ 14y ago
Does a long period of evolution actually imply it is "probably close to optimal"? I don't know a huge amount about evolution, but my two lines of logic would be: First, given the huge diversity of life on Earth, how can we argue that any species is "optimal". Secondly, the same argument could be made of life 500 million years ago, that it had been evolving for so long that it must be optimal, no?
- wzdd 14y ago
  It's actually a common misunderstanding of evolution to assume that it is optimising anything other than suitability to a particular environment (see http://en.wikipedia.org/wiki/Objections_to_evolution#Defining_evolution http://en.wikipedia.org/wiki/Objections_to_evolution#Definin...). So no, it doesn't make sense to use evolution as evidence that suitability in one domain implies suitability in another domain, especially when the domains have almost nothing in common except a shared requirement for management of scarce resources in a very abstract sense.
  - hnriot 14y ago
    I don't think you are seeing the parallels here. When you look closer you'll see that the domains are different but the problems are very similar and not so abstract, which explains the high degree of simularity and effectiveness of the congestion control algorithm. I would also take issue with what you see as a common misunderstanding. Evolution is, by it's very definition a process which adapts to a specific environment. Nothing more, no design, but optimal for any given environmental condition. When the environment changes rapidly, the pace of evolution increases, and vice versa. I don't think this is commonly misunderstood at all. Some people will always have no clue and of course there are those that think the fossils were planted by god, but aside from the out of band idiots, evolution is commonly correctly understood.
- rcthompson 14y ago
  For any living system that has had sufficient time to adapt to its environment, it can generally be assumed that the system has reached (or nearly reached) a local optimum for that environment, for the same reason that if you let a marble roll around on a surface for a long enough time, you can generally assume that it will come to rest at a local minimum. Of course, just as the local minimum of the marble is defined by the surface, the local optimum in the evolution of a trait is defined relative to the current environment. If the environment is rapidly changing, then you cannot assume near-optimality, just like you cannot assume the marble will come to rest on a constantly undulating surface.
  - 14y ago