Ending Feb 09
Feb 09 is not over yet, but I saw its end with my travel to Boston to give a talk. I had good conversations with Salil V, Michael M, David Parkes, Leslie V, and others. I spoke about internet ad auctions, so there were some in the audience from internet companies (besides the CS/Ecom/Sc types) and you can tell by the questions they ask. I realized, I really should write up open problems and directions that keep emerging.Science research tries to "model" a complex system by something simple, accurate, amenable to math and predictive. My hypothesis is that in complex internet systems (such as search, ads), humans change the underlying system faster than researchers can abstract and understand models, and hence, we should abandon the "model" based approach to research. For example, someone models Internet search users, but alas, an Engineer adds a new vertical search (say showtimes, products, celebrity profiles, whatever) and the models no longer apply. Diehard modelers will tell you this means we are just not modeling things at right granularity. And then they ask, what is the alternative? Experiment and tweak. As Leslie pointed out, I seem to be making a giant case for Machine Learning research. With a giant tweak that humans in the loop means humans alter the system, strategically.
ps: David Parkes, ever smart even over dinner, pushed me to a corner, saying experiment-and-tweak might just have us hike (climb?) the hills. Without models (inaccurate, non-predictive, mathematically amenable), we might never have new starting points.
Labels: aggregator
posted by metoo at 6:09 AM
5 Comments:
That's got to be one of the coolest "author bio" photos ever to grace a talk announcement....
They asked me for a photo about my work. I had nothing with plots, blinking lights, or plain hardware. It is an occupational deficiency of being an Algorithmer. So, I gave this photo, and many looked disappointed when they saw me in person, sans blondness. :)
-- metoo
Hi Muthu. I, like David, would have called you on your comment if we'd been at dinner. :)
I'm all for experimenting and tweaking. That's called "engineering", and it's a wonderful thing.
But modeling is also important. In the area you work on, your argument seems to be that we shouldn't get too involved in specific details in the modeling, since those are likely to change. But I think that's true in many modeling settings.
My analogy would be it's not so interesting to model the specific behavior of Facebook users on the current Facebook platform -- that could change tomorrow. But modeling social networks is a fascinating field. There's something more general there that provides important high-level lessons (and occasionally detailed results).
There are cases where the baseline is so entrenched it's good to model down to fine levels of detail and aim to obtain precise results. In other cases, models are obvious approximations where you're trying to just learn rules of thumb. (Are there important phase transitions? Do things typically enter a stable state or is the system fundamentally unstable?)
As another example, CS people have often dismissed large parts of queueing theory because it often assumes Poisson process inputs. But if you think the point is not to extract exact numbers but rule-of-thumb behaviors, this approach can be just fine -- giving some important insights. So too, modeling can provide important insights even when the engineering is moving at Internet speeds.
Thanks for coming to visit and the entertaining talk!
Hi Michael,
I knew I had to have a separate post for things we discussed (make it, argued, :)) and for things we did not have chance to argue but I knew we would (eg., the YMGA problem). Now I realize we would have argued about more things! Thank you, thank you for the great time.
Modeling has the feel of explaining the observations years after the world came to be, not often helpful in predicting things (did social or web graph model in vogue while ago actually help us predict anything?). In some worlds such as the Internet systems (not all worlds) where humans have resources to process data and understand them, they also have the resources to change the system rapidly: will the modeling approach work and help predict? This is essentially a question inspired by Engineering, and one of the intended outcomes of research.
- metoo
where humans have resources to process data and understand them, they also have the resources to change the system rapidly:
This seems akin to control theory, where one has to come up with a model that incorporates the fact that the system changes on the basis of the input given to it.
Post a Comment
<< Home