To follow up - page 45

 
Mathemat >>:

не могу понять, как предыдущее Обучение-II может превратиться в приготовление к психическому расстройству. Это просто обучение "способу видеть кляксу" (см. ниже), который в системе ценностей мага не может быть противоречивым.

It seems to me that the translation there is not very good, maybe the author meant that madness is impossible without Learning-II.

I remember reading a very long time ago a fantasy story, where they drove mad by alternating coloured spots on the screen :). If the magic has been turned off for a long time and the mage continues to make passes, he will look at least not quite adequate from the outside.

 
Candid >>: А вообще, если магия давно отключилась а маг продолжает делать пассы, он будет выглядеть как минимум не совсем адекватным со стороны.

Well, that's how 99% of magicians work :) They don't get shit done, but they put a good face on a bad game in order to get money out of the client. And at the same time and hang a lot of noodles on their ears.

And in general, it seems that the vast majority of people never go beyond Learning-II in their lives (again, because of the self-affirming punctuation). Moreover, it is not welcome at all, as it is a dangerous thing to do.

Well, that's all right. The most interesting thing is the CU classification. Stupid two-machine system without any clever context - is it zero learning or learning-I?

It's somehow easier with people. Learning-III is in any case at least temporary insanity, even if one loses the ability to change it after a life paradigm shift.

 

Mathemat : Ай-яй-яй, как это знакомо. С другой стороны, не могу понять, как предыдущее Обучение-II может превратиться в приготовление к психическому расстройству.

I've been thinking for a long time in between about what to say to you since last night. I started to formulate something... but then I changed my mind. It's going to be long and tedious, kind of off-topic for the forum.

Here's a link to a page where you can download Bateson's book Ecology of the Mind. There (in the book) you'll find the article "Minimum requirements for a theory of schizophrenia".

I recommend to start your analysis with this one - it's a good article on the subject. // comprehensible even to psychiatrists, for whom this lecture was read in 1959 ;-)

And very interesting, though a bit sad. At its core. But... what can you do, that's the topic.

// And on the downloadable page itself there's a charming "Metalogies" from the same book - you'll like them - take a look in passing.

For now, let's continue to deal with TCs, and if it's interesting to look into this phenomenology in life, we'll come back and look into it.

I hope the topic starter won't mind too much, philosophical deviations are kind of welcome.

 
Mathemat >>:

1) Ну так 99% магов и работают :) Ни хрена не получается, но делают хорошую мину при плохой игре, чтобы вытянуть бабло с клиента. И при этом еще и вешают кучу лапши на уши.

2) Ну это ладно. Самое интересное - это классификация ТС. Тупая система на двух машках без всяких умных контекстов - это нулевое обучение или Обучение-I?

3) С людьми все как-то проще. Обучение-III - это в любом случае хотя бы временное помешательство, даже если человек после сдвига жизненной парадигмы теряет способность ее менять.

1) In general, it makes sense to take a deeper look at "magical thinking". The everyday variety is superstition. Although it has long been known that it is not good to believe in omens (there is such an omen),

but people aren't really retrained to live without superstition. Take this forum for example. For example, here's a neighbouring thread at.... Okay, I'm not gonna point the finger. I'm not supposed to have a horoscope. :)

2) Stupid two-machine system is a system in which there is a rigid connection between input and output that does not change over time. There is sort of no learning (0).

But we can look at the system from this angle - a programmer taught some design to react by trade transaction to a certain ratio of strokes.

Then we kind of have a training-I. Here we can agree on how to look at this matter. I would personally suggest the second way.

That is, we see the program as (a) at the beginning not able to do anything (b) trained by some stimulus-reactive activity (kind of like a programmer :).

This is not a strict (essentially incorrect) use of Bateson's terminology, but it seems convenient to me. As long as we don't forget that initial incorrectness,

I don't think there will be any victims. :-) The floor is open to the people. What do you prefer?

3) :))

 

There is a primary reality - the stream of quotes. The first stage of its subjectivisation is the overlay of mash-ups. This is a basic, primary partitioning into alternatives within the limits of one rigidly defined set of them. But one can impose RSI instead of waves. It would be a different division. There will be another living being with a qualitatively different basic set of primary alternatives.

Обучение-I есть изменение специфичности отклика благодаря исправлению ошибок выбора внутри данного набора альтернатив.

In short, let the dumb "two wipes" be Learning-I. OK, now put contexts (CCs) on top. We get something similar to O-II already. However, our set of primary alternatives itself hasn't changed, as it is rigidly set by the same mashkeys. The living thing remains the same: we have the same rat. Theoretically, we could try to translate the rat into O-III, but it won't work: the set of primary alternatives is too poor for that. In order to get a living creature capable of O-III, you need to change that basic set (instead of mashka - ZZ, channels or Phoebe, for example). Yeah, I think, even for O-II, our rat ("two mashka") is not too great ability: it already at this level will go crazy...

What is the point of all this? The point is that you can argue endlessly about 'ideal CCs', but their possibilities will be fundamentally limited by a set of primary alternatives, i.e. the 'class of the living being'.

Not too much of a load yet?

Accordingly, the next problem arises: we have to learn to figure out in advance which set of primary alternatives (mashek-based partitioning, ZZ, Fib or whatever) has the richer ability to, well, at least to O-II.

By the way, I have wondered about this question before, but a reasonable context for its posing only arose thanks to this article by Bateson. Thank you, Volodya.

 
Mathemat >>:
.....

Кстати, этим вопросом я задался и раньше, но разумный контекст его постановки возник только благодаря этой статье Бейтсона. Спасибо, Володя.

You are welcome. The article really is a good compass in a sea of learning processes. I've never seen a better one.

// I'll answer the rest later, I need to think about it.

 
Mathemat >>:

Есть первичная реальность - поток котировок. Первый этап ее субъективизации - это наложение машек. Это базовое, первичное разбиение на альтернативы в пределах одного жестко заданного их набора. Но можно ведь и RSI наложить вместо машек. Будет другое разбиение. Будет другое живое существо, у которого базовый набор первичных альтернатив качественно другой.

Короче, пусть тупая "две машки" будет Обучением-I. ОК, теперь накладываем сверху контексты (КК). Получается уже что-то похожее на О-II. Однако сам набор первичных альтернатив у нас не изменился, т.к. его жестко задают те же машки. Живое существо осталось таким же - та же крыса. Теоретически крысу можно попытаться перевести на О-III, но не получится: слишком бедный для этого набор первичных альтернатив. Чтобы получилось живое существо, способное к О-III, надо сменить этот базовый набор (вместо машек - ЗЗ, каналы или Фибы, к примеру). Да, думаю, даже и к О-II у нашей крысы ("две машки") не слишком большие способности: она уже на этом уровне с ума спятит...

К чему это все? К тому, что можно бесконечно рассуждать об "идеальных КК", но их возможности будут принципиально ограничены набором первичных альтернатив, т.е. "классом живого существа".

Еще не слишком сильно загрузил?

I have an easier way of looking at it. You may not have the article settled yet.

Let us teach MTS to play with the following algorithm (I'm replacing your two moves with one of my MACD):

1) if the MACD crosses the zero line upwards - close all bets downwards and open one lot upwards

(i.e., call the UP() procedure which does everything necessary to make us stand one lot up)

2) if the MACD crosses the zero line downwards - close all bets upwards and open with one lot downwards. (i.e. call the Down() procedure).

We have a system of O-1. // One stimulus (MACD zero crossing) - one reaction (trade) // (more precisely 2 stimuli, two reactions.)

Now let's add contexts. For example (1) = Abs(RSI-50)>30 (2) = Abs(RSI-50)<=30

Contextualize, for example, by "hard" partitioning into classes (Boolean function).

Accordingly, the algorithm becomes more complicated. For example:

if (Abs(RSI-50)>30)

{ if (MACD()>0) { UP(); } else { DOWN(); } }

else

{ CloseAll(); }

Now we have the O-2 system.

If we change the set of contexts in this algorithm, we are thereby overtraining-2.

If we change the reactions, keeping the context set the same, we also overlearn-2.

If we change both, that will also be retraining 2.

Finally, if we create a whole branching hierarchy of contexts and equip each terminal context with a whole table of single-valued reactions to various stimuli (as elementary as, for example, reactions to crossing of a waving bag), we would still only be doing retraining-2.

I.e. O-3 is something qualitatively different, it is an ability of the system to correct its own learning-2.

I would like to finally build such a toy, I don't see anything "inconceivable" in it, although it's not easy.

Actually, this is what we can do after building good models-2 - i.e. incapable of self-training but always ready to crawl into external optimiser and optimise. :)

Accordingly, the following problem: we must learn to estimate in advance which set of primary alternatives (Mashki, ZZ, Phoebe, or whatever) has richer abilities well at least to O-II.

We should. I don't mind. I'm learning a little bit. There is even a work in progress. I'm a little stingy, but I might post something later. Let's get used to the terminology and general logical considerations, and work together on the possibilities. And the difficulties, too.

So far, considerations are as follows: the simplest classification schemes (such as the one in Sorento, or the structural equivalent in this post) have minimal potential.

They are able to "fit" a piece of the past, and leave hope for the short-term persistence of the pattern grasped.

// I call such things "primary oracles" :) That is, sort of have the right to the name, yet for some time after optimization, as a rule, the MO gain from them > 50%.

Not bad in principle. Considering that such a design is not difficult to automate and make it self-optimizing in a sliding window adjacent to the NV (NV - Present Time).

Practice with the optimizer shows that the chances of such a design are not bad.

But it is possible to go further. Branch contexts, build more sophisticated ones from such "primary oracles", etc. etc.

I look forward to further considerations.

 
MetaDriver >>:

Я как-то проще смотрю на это. У тебя возможно не устаканилась ещё толком статья.

How else would she have settled down if I only saw her yesterday :) It's just a bit of a fantasy.

I'm waiting for further thoughts.

Yeah, I'm thinking about it. The article needs more reading.

 
the simplest classification schemes have minimal potential.

Они способны "подгоняться" под кусок прошлого, и оставляют надежду на кратковременное сохранение схваченной закономерности.

..

But we can go further. We can branch contexts, build more sophisticated ones from such "primary oracles", etc. etc.

Awaiting further thoughts.

Right. But what nobody wants to discuss is the parameters of the context....

Looking forward to further considerations.
 

Sorento, this thread has perhaps long since become an ideological-philosophical thread. Here it is probably better to discuss "broad strokes" - what defines "trends", i.e. fashion. It was not MetaDriver or myself who set the direction of this thread, but the Weissmanist-morganists (although at the beginning it was highly specialized).

I think it makes sense to search for the principles of search and selection of AC parameters first in order to specify them later (probably better in a specially created branch). I don't have any ideas yet, how I don't have any idea how to find them yet, but I hope they will appear. If you already have these principles in mind, why not discuss them?

I think it is logical to look for context parameters in close relation to a set of primary alternatives ("class of living being"). I do not like a chaotic and haphazard enumeration of these parameters in the hope that "someday they will somehow come together". Futility of this approach is obvious to me when I see another super-system made of, say, MACD, Bollinger, Stochastic and channels with parameters adapted in an unknown way.