[p2p-research] Drone hacking

[J. Andrew Rogers]

On Tue, Dec 22, 2009 at 3:56 AM, Andy Robinson <ldxar1 at gmail.com> wrote:
> And given that they only have data from massified processes, their data is
> always incomplete.  Since the context if complex and chaotic, this could
> easily be fatal: the missing piece of data, however apparently trivial,
> could make the difference in predicting how someone would actually act.


Real-world data very rarely has anything approaching cryptographic
lack of predictability. Chaotic systems are not unpredictable per se,
just divergent. If you build the model using 27 different measures,
you have a good probability that most of them will be well-behaved for
a particular prediction. You don't need perfection, just "good
enough".


> Hence, the calculations in question will only be useful for
> predicting/directing massified behaviour, and only for as long as people
> remain massified and their lives reproduce familiar patterns which have
> already been observed a great many times.


These kinds of methods are robust because the patterns are high-order
and non-axiomatic. You can use the high-order patterns to predict
low-order behavior in new contexts.


> Actually I would like to see real
> information rather than assertions here - what proportion of the time do
> these 'primitive' versions of the coming supercomputers actually
> successfully predict even something as simple as online buying behaviour?


Using data limited primarily to behavior on a large website, which
lacks in-depth context, behavior prediction buys an individual sales
increase in the low tens of percentage points. Call it 10-30% as a
rule of thumb. There are secondary markets for these customer
behavioral models, since few businesses have the wherewithal to do it
themselves. Performance metrics are trade secrets.

Performance has steadily improved over time but is still limited by
practical technical factors.


-- 
J. Andrew Rogers
realityminer.blogspot.com