On Mar 18, 6:19 am, "Ted Dunning" wrote:
> On Mar 15, 3:37 am, jones...@emporia.edu wrote:
>
> > In conventional capitalist economics one assumes value
> > monism and a scalar utility
>
> I really don't think that the convenience of having a scalar objective
> function is philosophically related to capitalism.
It's not a question of convenience. If values are components
of a value vector and utility is a vector then scalar
utility (money?) is simply incorrect, and capitalism
along with it.
The reason for
> optimizing a scalar-valued objective is that it is a much simpler
> problem to state.
But it isn't simple. How many dollars is a human life
worth? Or your wife's love? Or your vote in the next
election? Or the not-guilty verdict in your trial? Value
monists have more than "difficulty" in making a scalar
utility work.
Real numbers form a total order which lets the
> definition of "optimum" make sense.
Perhaps life is not about optimums.
>
> If the parameters are real valued then you also get things like
> gradients and Hessians to work with.
>
> > It may be, however, that value monism is wrong and we
> > can not reduce all rewards to a single scalar
>
> As you note, this idea is hardly new.
I don't think I was making any claims about newness
anywhere in my post. If asked I guess I'd say what
was new is that I have an AI that I can flip back
and forth from being a value monist to be a value
pluralist. And the behaviors are different for each.
>
> But you should observe that most approaches to multiple objective
> optimization start with some way of defining an order function on R^n,
> if only so that you can say what "optimum" means. In many approaches,
> this is done by defining some heuristic way of reducing the multiple
> cost functions to a single value.
>
> > Initial results suggest that use of a vector utility
> > may make the system less likely of get stuck in local
> > maxima. If the system can not improve L, for example,
> > it may be able to increase N. After evolving for a while
> > one may then find L and N can both improve.
>
> This statement is way too over-arching.
I made the title of my post a question. I would just
be interested to know if value pluralism had an
evolutionary advantage over value monism. And yes, it
would have consequences for capitalism.
>
> First, there are many methods that avoid the problem of local minima
> using scalar objectives.
I don't doubt there are other ways to avoid or get out
of local maxima. But it would be interesting to know
if value pluralism had some advantage over value monism.
Something that made the value pluralist more fit.
These include simulated annealing various
> evolutionary algorithms. There are also a number of methods in many
> settings that avoid the need for optimization, per se. For instance,
> maximum likelihood methods are often better replaced by sampling
> techniques which evaluate posterior likelihoods directly. Thus,
> scalar objective functions with local minima are not a hopeless
> problem.
>
> Secondly, it has been known for some time that the introduction of
> auxiliary variables can make many problems much easier.
It may be that value pluralism is working in just that way in
Asa H.
Latent
> variable techniques are an impressive example of exactly this. For
> instance, in regression problems where you are trying to estimate a
> few parameters, it can be helpful to introduce two latent variables
> per observed data point. Thus, this idea that increasing the
> dimensionality of the problem is useful is hardly novel.
>
> Most importantly, though, all of this has nothing to do with
> capitalism or alternatives.
I don't agree but that was not the subject of my post.
>
> But you may also have something interesting.
>
> Can you say more about your specific examples and methods?
I have emailed you a copy of my Asa H paper. I also refer
you to my other references.
>
[ comp.ai is moderated ... your article may take a while to appear. ]
--- SoupGate-Win32 v1.05
* Origin: you cannot sedate... all the things you hate (1:229/2)
|
