We monitor cell phone communications coming from certain parts of the world with the hope of detecting terrorist plans or intentions. They learn of this so they switch to other forms of communications. The question this raises is an old one.
How does or can the act of observing affect the behavior of the things being observed?
This question can be re-framed with regard to decision support system or decision analysis in which case the question becomes, in a much more indirect sense, how are the predictions inherent in generating data presented to users to aid in decision support affected by the ultimate decisions that are made based on them as those same predictions and decisions are made repeatedly over time?
In simpler terms, when does, if it does, the tail begin to wag the dog?
If it does, then how can such decision support systems compensate or dynamically adapt such that predictions remain valid and in fact become more accurate over time?
5 comments:
The question of how observation affects the behavior of the thing being observed is a very interesting one. Neal Stephenson's novel "The Cryptonomicon" provides a very well thought out (and highly entertaining) treatment of this question and I would highly recommend this novel (my favorite). The protagonist is a young mathematical genius who is working to break the secret codes of Japan and Germany in WWII. He comes to the conclusion that the allied forces must be very careful how they use the decoded messages, or the Germans and Japanese will be able to infer that their codes have been broken and will be able to act accordingly.
Jeff
http://jeffreyellis.org/blog/
Thanks for the recommendation.
I would also be interested in works that discuss how our response to predictions or estimations made by a system, be they statistical, qualitative, quantitative, or otherwise, affect subsequent predictions and estimations from that same system? How can/do these systems correct for such feedback?
I want to make sure I understand your question. I think this is all in the context of a closed-loop system, whereby our response to a system output (e.g., a decision support system's presented data) will change subsequent system outputs (e.g., the system will have to take into account how we responded to its previously presented data). Do I understand the question correctly?
If so, I do not know of any works that explicitly discuss this on a macro scale. Although, my master's thesis may be somewhat related (see http://www.geocities.com/jeff_ellis.geo/thesis/thesis_front.html). In my thesis I investigate the use of a predictive simulation to overcome the problem of remote controlling a system in the presence of communication time delays.
On a micro scale, this question reminds me of quantum mechanics, where the very act of observing something can change the behavior of the thing you are observing. This happens because quantum-level particles exist in an ill-defined probabilistic "quantum state" until the act of observing forces it to choose a specific state. Probably not relevant here, though.
Jeff
http://jeffreyellis.org/blog/
Yes, I am thinking about macro scale but sometimes solutions transcend scale.
The type of contexts I am thinking of would not necessarily be completely closed though. Systems that involve informing users about resource allocation optimization potentials through a simulation where probabilities are incorporated into what-if types of analysis. Where actions are taken in the system being simulated based on the analysis which then changes the simulation because it is being feed by data from the real system and so on.
It turns out that the situation you described is really closed loop although not explicitly so. Another example would be if I had an advanced simulation of the stock market that I used to predict the performance of a certain set of stocks, and then gave that info to a group of high-powered investors. Their actions based on that info will change the value of those stocks (they will drive the prices up/down with their buying/selling) and render my predictions wrong. What would be nice here is if my simulation could take their future actions into account, and come up with the predictions that, when provided to the investors, will result in that prediction being correct. Very interesting problem, it seems like only an iterative approach would work here and one that also simulates the target audience's behavior.
Post a Comment