- --- In ai-philosophy@yahoogroups.com, "Eray Ozkural" <erayo@...>
> That is, I am hoping that it is completely obvious to you that you can

For the umpty-umpth time: probabilites are possibilities with numbers

> have a probabilistic point of view that doesn't mention any iffy

> entities like possible worlds.

attached. And you can talk about PW's without reifying them. - --- In ai-philosophy@yahoogroups.com, "jrstern" <jrstern@...> wrote:
>

no

> The strong point of a probabilistic approach is that you need have

> knowledge of any causal chain from A to Z.

skipping...

> To this day, I don't know what to do with any of this. If it works,

it

> works, but I see it as an absence of strong theories, I cannot see

the

> statistical approach as foundational.

Well, I suppose if you do have the statistics, then you have a

foundation :)

I do not want to underestimate any controversies on the

interpretation of statistics but IMHO the promising thing about

statistical inference is the recently developed framework of

graphical models which provides a mathematically rigorous way to

model causality relations among some entities and couple the local

knowledge on them (in terms of distributions) constructing the global.

By themselves, they yield expert systems; automatons that perform

some specified tasks. But through LEGO type split/merges led by an

appropriate cognitive architecture I think bigger problems could be

solved.

Lack of strong theories is a problem for proposing such an

architecture I think; not for statistical inference itself neither

for choosing it as the foundation...

Murat UNEY - --- In ai-philosophy@yahoogroups.com, "zeb_6662001" <e120532@...>

wrote:>

Not at all. You have a behaviorist model, and the traditional

> --- In ai-philosophy@yahoogroups.com, "jrstern" <jrstern@> wrote:

> >

> > The strong point of a probabilistic approach is that you need

> > have no knowledge of any causal chain from A to Z.

> skipping...

> > To this day, I don't know what to do with any of this. If it

> > works, it works, but I see it as an absence of strong theories,

> > I cannot see the statistical approach as foundational.

>

> Well, I suppose if you do have the statistics, then you have a

> foundation :)

problems of determining whether the categories for your model even

make sense. The virtue you suggest is that there is no bright line

between analytic and synthetic, schema and data, therefore any

combination that works, works. But it is completely non-foundational

in the way that it works.

> I do not want to underestimate any controversies on the

global.

> interpretation of statistics but IMHO the promising thing about

> statistical inference is the recently developed framework of

> graphical models which provides a mathematically rigorous way to

> model causality relations among some entities and couple the local

> knowledge on them (in terms of distributions) constructing the

>

Well hey, it's not that humans are the most accurate possible

> By themselves, they yield expert systems; automatons that perform

> some specified tasks. But through LEGO type split/merges led by an

> appropriate cognitive architecture I think bigger problems could be

> solved.

>

> Lack of strong theories is a problem for proposing such an

> architecture I think; not for statistical inference itself neither

> for choosing it as the foundation...

predictive agents. Back in expert system days, I heard it repeatedly

stated that simple statistical regressions were more accurate than

most expert systems, and far, far easier to construct. In fact, such

quantitative systems were often more accurate than the human experts,

it was said, much to the shock of the human experts (and more

especially the human non-experts) involved.

Might be true, for a large class of systems. Wouldn't surprise or

shock me. But it's not AI, or cognition, or anything of the sort.

AI is NOT about computing the right answer, or a simple calculator

would rate very highly indeed, and we would have to rate the pool

table as most "intelligent" about working out the answers to various

mechanical problems. It is accurate, sure, in fact it is

determinative, but that does not make it "intelligent".

It's a standard argument. For simplicity, I take the instrumentalist

view that algorithm X is best called AI whatever the results, and

algorithm Y is best NOT called AI, whatever the results.

Josh - --- In ai-philosophy@yahoogroups.com, "jrstern" <jrstern@...> wrote:
>

...

> Well hey, it's not that humans are the most accurate possible

> predictive agents.

> But it's not AI, or cognition, or anything of the sort.

...

> AI is NOT about computing the right answer,

I agree and indeed what I was trying to tell is that an architecture

that would play with those models would lead to AI or something else

(Non-AI but Expert System, non-AI but a fancy automaton). My point is

that a probability space is fine enough to start with (in the simplest

sense, given two random variables we are equiped with nice concepts

revealing their underlying connection). Moreover, these relations are

modular in some sense when we introduce more variables and some

relations with the previous ones.

The term Statistical inference should not be misleading that it would

result the most accurate answer. It would lead an approximate but

computable answer depending on the model of the algorithm and the

physical entities involved.

The thing is that probabilistic modeling seem to provide kind of

modular modeling tools which would be used as bricks by higher level

constructs. If these constructs merge and build more complex ones, they

would still show more or the less same properties of the brick models;

some sort of scalability and one of the following; accurate results,

approximate results in the long run or totally misleading results.

These are some nice properties which I expect from an architecture - a

meta program structure which would be labeled as an AI. That' s why I

ironically pronounced them "foundational", not because they would

provide accuracy but modularity and scalability.

We talk about probabilities for the umptiest humptiest times because we

seem to "naturally" morph the concepts on AI that have been mentioned

here more than umptiest humptiest times, to those in the theory of

probabilities.

I do think that AI is a matter of architecture...

MU - On 7/10/07, zeb_6662001 <e120532@...> wrote:

> The thing is that probabilistic modeling seem to provide kind of

Right. This is also what I meant when I said that a probabilistic

> modular modeling tools which would be used as bricks by higher level

> constructs. If these constructs merge and build more complex ones, they

> would still show more or the less same properties of the brick models;

> some sort of scalability and one of the following; accurate results,

> approximate results in the long run or totally misleading results.

framework might be able to combine together smaller machines,

which we already have: classifiers, function inference machines,

logical planners etc. I don't know if we can tie together such

a variety of machines, but why not?

Best,

--

Eray Ozkural, PhD candidate. Comp. Sci. Dept., Bilkent University, Ankara

http://www.cs.bilkent.edu.tr/~erayo Malfunct: http://myspace.com/malfunct

ai-philosophy: http://groups.yahoo.com/group/ai-philosophy