> I think I was pretty clear that my goal was to "make everyone happier"
How does your behavior achieve your goal, then? How do you know what would make people happy if all of your information about what makes them happy is snapshots of the past? Maybe a better study of the model would yield that the best way to make people happy is to offer them the exact opposite of what they had in the past? I don't understand how you can draw conclusions from data without understanding the model the data is a snapshot of.
> It's an irrelevant one, however, since you can't actually predict whether any action will change this distribution.
I didn't say that. I said that we can't predict with certainty and we're not trying to make predictions at all, but we do know from history that political and social action does work. That's how women were allowed to vote, go to universities, and become doctors and lawyers.
> that result in no testable predictions, I'm sorry for you
As I have spent even more years studying mathematics, I can tell you that neither subject was a waste of time, testable predictions or no. I feel sorry for you if testable predictions are the only measure of worth you know, but I hope that you are at least aware that much of the progress humankind has made wasn't due to any scientifically-predictable actions (and some of it was).
Nevertheless, I never said I can't make any testable predictions, only that we can't predict the course of history.
> Perhaps it might have been a better use of time to spend a semester studying complex systems instead.
Well, I did both.
> but the nonexistence of good predictions does not follow this - it merely changes the character of the predictions.
Really? If your measurement time (of initial conditions) significantly exceeds the time the predicted trajectory is within usable error, then yes -- nonexistence of predictions does follow. That doesn't mean, however, that we've seen that societies where social action was taken changed faster than those that didn't have social action. If nonlinear equations is your field then you know that, sadly, sometimes we are reduced to making qualitative predictions only.
No one disputes the necessity of models. Why do you think I did?
I didn't say that. I said that we can't predict with certainty...
Actually you said "I cannot say that my predictions are better than random". Either your predictions convey information (i.e., posterior(world state|prediction) != posterior(world state)) or they doesn't. If they don't, they are useless for prediction since E[utility|choice] cannot vary.
If they do, you can make the basket of predictions I asked for and behave better than random chance.
Really? If your measurement time (of initial conditions) significantly exceeds the time the predicted trajectory is within usable error, then yes -- nonexistence of predictions does follow.
The nonexistence of stable predictions of x(t) does indeed follow. Again, there are a huge number of quantitative predictions which can be made. For example, many systems with positive liapunov exponent allow claims like "a particle's long term distribution will behave like a sample from the distribution f(x) dx" or "a particle with energy < E_0 will travel a distance L only with probability < A exp(-BL)."
Because you said that ignoring people's past preference leads to sub-optimal decisions and implied that the best decision is merely to reflect past actions. You can't possibly make that claim unless you have a model that states that people's preferences don't change. That would be a very wrong model, though, because we know that they change and we know what may change them.
> Actually you said "I cannot say that my predictions are better than random".
By that I meant predictions of the course of history. I can say with high certainty that social movements often achieve their goals, and without them, those goals arrive much, much later. In particular, the feminist movement (on all its forms) has had tremendous success since the middle of the nineteenth century, and there is little reason to believe that the trend will stop. OTOH, the resistance to change has also taken a familiar path (while historians don't try to predict the future, they do recognize certain patterns) of an unofficial glass ceiling, then appeals to "nature" and "science", a better past and the tyranny of the reformer against a sieged hegemony. All of these are very familiar to us.
> a particle's long term distribution will behave like a sample from the distribution f(x) dx
Sure, we could theoretically make that claim, but we wouldn't be able to verify it because our measurement time exceeds the duration of any reasonable error of the prediction.
Listen, I went into history after studying math (yes, including non-linear equations), with delusions of being the first to come up with a dynamical model of society's behavior (with "power" of course, being a central quantity). But then I learned history and realized how unattainable that goal is. Maybe one day, if we have near instantaneous means of measurements. Otherwise, coming up with a model out of grossly inaccurate measurements made at different times -- and far too few of them (there are so many variables, so many forces, including forces of nature, namely a lot of external energy) -- all mediated by narrators, makes it impossible. We can say that "social movements that grow large enough tend to work", but putting that into any formula won't make the prediction any more accurate. I still have a dream of a semi-accurate model of how power flows in society, but it will require much better measurements.
BTW, if all our models are no better than random, the basis for relying on past behavior as a predictor of future happiness is no less shaky than any other. In that case, we'd have no reason to believe that ignoring past data would lead to any worse decisions than any other assumption.
How does your behavior achieve your goal, then? How do you know what would make people happy if all of your information about what makes them happy is snapshots of the past? Maybe a better study of the model would yield that the best way to make people happy is to offer them the exact opposite of what they had in the past? I don't understand how you can draw conclusions from data without understanding the model the data is a snapshot of.
> It's an irrelevant one, however, since you can't actually predict whether any action will change this distribution.
I didn't say that. I said that we can't predict with certainty and we're not trying to make predictions at all, but we do know from history that political and social action does work. That's how women were allowed to vote, go to universities, and become doctors and lawyers.
> that result in no testable predictions, I'm sorry for you
As I have spent even more years studying mathematics, I can tell you that neither subject was a waste of time, testable predictions or no. I feel sorry for you if testable predictions are the only measure of worth you know, but I hope that you are at least aware that much of the progress humankind has made wasn't due to any scientifically-predictable actions (and some of it was).
Nevertheless, I never said I can't make any testable predictions, only that we can't predict the course of history.
> Perhaps it might have been a better use of time to spend a semester studying complex systems instead.
Well, I did both.
> but the nonexistence of good predictions does not follow this - it merely changes the character of the predictions.
Really? If your measurement time (of initial conditions) significantly exceeds the time the predicted trajectory is within usable error, then yes -- nonexistence of predictions does follow. That doesn't mean, however, that we've seen that societies where social action was taken changed faster than those that didn't have social action. If nonlinear equations is your field then you know that, sadly, sometimes we are reduced to making qualitative predictions only.