To follow up - page 46
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Я считаю, что параметры контекста логично искать в тесной привязке к набору первичных альтернатив ("классу живого существа"). Мне не нравится хаотичный и бессистемный перебор этих параметров в надежде на то, что "когда-нибудь они как-нибудь срастутся". Бесперспективность такого подхода очевидна для меня, когда вижу очередную суперсистему, собранную, скажем, из MACD, Боллинджера, Стохастика и каналов с непонятно как подогнанными параметрами.
Exactly... And you and I are not the only ones to think so.
Sorento : maybe we should forget classical indicators for a moment and ponder - what properties, in principle, characterise CR in terms of its movement?
But philosophising endlessly is not pragmatic. We have to start moving.
The assault will run out of steam faster if we don't conquer a single height and get a tiny, but practical result...
Mashki with 33 shocked me a bit.
Let's pick apart the bones of this 'homework'. especially as anyone can get the data.
As I understand it, the source of the breakdown was the bimodality of the ideal response distribution...
and it has been resolutely eliminated.
If such an algorithm works on a primitive level, why not try to apply it more widely?
2) Тупая система на двух машках - система в которую прошита жёсткая связь между входом и выходом, которая со временем не изменяется. Вроде как обучение отсутствует (0).
Но можно на эту систему взглянуть под таким углом - программист научил некую конструкцию реагировать торговой транзакцией на определённое соотношение положений машек.
Тогда вроде как имеем обучение-I. Здесь можно договориться как на это дело смотреть. Я бы лично предложил вторым способом.
Т.е. рассматриваем программу как (а) вначале ничего не умеющую (b) обучившуюся некой стимул-реактивной деятельности (вроде как у программиста :).
Это не строгое (в сущности неправильное) использование терминологии Бейтсона, однако мне оно представляется удобным.
Well full compliance with Bateson's terminology is not our goal.I would draw attention to his words that a hierarchy of training corresponds to a hierarchy of contexts. In relation to a mechanical trading system, classification in the language of contexts (a hierarchy of contexts) may be more adequate and less confusing to the mind.
In general, it is possible to take a creative approach and create a biocenosis of sorts. The lowest level of food chain is primitive entities with purely reflexive reactions, at least to the same mashka. Some kind of bacterius speculatis :), however, Vladimir is an expert on the terms :)
The next levels must feed on the previous ones. The creation top should be able to determine which particular timeframe should be used at the moment and which direction to take in case of crossing. The higher the level, the smaller the span of population waves should be.
By the way, one can consider returns as such entities, then the next level is inputs by this or that oscillator calculated on the basis of returns.
I think it would be a mistake to try to decide in advance which primary context (markup) has greater prospects. Let a hundred colours bloom. It's quite possible that there really is no fundamental difference (in terms of trading results). Although I personally think that the fractal nature of the market corresponds to ZZ to the greatest extent.
------------------------------------
Т.е. О-3 - это уже нечто качественно другое, это способность системы самостоятельно корректировать своё обучение-2.
Такую игрушку и хотелось бы в итоге построить, ничего "немыслимого" в этом не вижу, хотя это и не просто.
Собсно этим можно и заняться, после построения хороших моделей-2 - т.е. не способных к самостоятельному переобучению, но всегда готовых залезть во внешний оптимизатор и пооптимизироваться. :)
Only systems capable of self-learning (not auto-optimisation) that do not use templated blanks such as waving crosses would probably fall under these concepts. Autochanging waving parameters would be auto-optimisation. (highlighted in blue).
Evolutionary development of model-2 research cannot move to model-3 (highlighted in yellow), because one would still have to abandon the template concepts of model-2.
Either deal immediately with models-3, or not deal at all, as it is impossible to come to them "gradually". IMHO.
Finally, if we create an entire branching hierarchy of contexts and equip each terminal context with an entire table of unambiguous reactions to various stimuli (as elementary as, for example, reactions to crossing waving), we would still only be overlearning-2.
I.e. O-3 is something qualitatively different, it is the ability of the system to correct its own learning-2.
I would like to build such a toy as a result, I see nothing "inconceivable" in it, although it is not easy.
Actually, this could be done after building good models-2, i.e. incapable of self-training, but always ready to crawl into external optimizer and optimize itself. :)
Utopia. I mean highlighted.
1. This is not training your EA at all, much less training 2.
2) This is essentially your own training-2, which you are going to formalize in a relatively rigid scheme and put it into Expert Advisor. However, do you have the ability to do this Bateson training-2 ?
3. to "create a whole branching hierarchy of contexts", and with "a whole table of single-valued reactions" to each of them, one must have a method, a tool, a way to identify these contexts, a way to determine the effective reactions. Do you have it? What is it? What is it based on? Where did it come from? If you have it, you can safely put this article on the shelf. It is enough to create a successful EA without "learning". But the problem is that you don't have it, and you will not be able to come up with it out of your head. And an article will not help here.
As for "not capable of self-training, but always ready to get into an external optimizer and optimize" programs, this goodness is neither rare nor targeted. And Bateson has nothing to do with it.
So what's the bottom line ?
And the bottom line is that learning has as its goal the right behaviour. And correct behaviour is the choice of correct reactions to external circumstances. Our reactions are fixed: buy, sell, smoke. So there is only one thing left to do: assess the circumstances adequately. And we come back to
Accordingly, the following problem arises: we must learn to figure out in advance which set of primary alternatives (breakdown on the basis of mash, ZZ, Fib or whatever) has the richer ability to well at least to O-II.
I think it makes sense to first look up the principles for finding and selecting QC parameters to further flesh them out (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's logical to look for context parameters closely tied to a set of primary alternatives ("class of living being"). I don't like the 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.
So we again return to the problem of parameterization. However, it does not stand alone. And here I completely agree with Alexey.
Parametrization is a consequence of the model and not a set of numbers plucked at random. Parametrization is also a property of the model. Within a model it cannot be changed arbitrarily. And the success of parameterisation is entirely determined by the adequacy of the model.
Why do we keep coming back to MACD, stochastic and other stuff? They are just odd numbers and don't make much sense. Can anyone suggest a model, in which they would play at least a reasonable role? If not, why talk about them?
is again forced to mentally return to the two ideal characters, the indifferent observer and the yapping dog.
To the older and current zigzag.
To the two wipers with periods of 15 and 200.
-----
Analogies are very simple, and it's not for nothing that wipers are "fast" and "slow".
They characterise the behaviour of the price. Is it bad?
Suggest what's best...
It doesn't seem to work out with a "home building" SCO.
;).
Правильно. Но вот параметры контекста никто обсуждать не хочет...
He wants to. But is shy (like you). Or the toad has strangled him (like me). ;)
Sorento, эта ветка, пожалуй, давно превратилась в идейно-философскую. Здесь, вероятно, лучше обсуждать "широкие мазки" - то, что определяет "тренды", т.е. моду.
(1) Сложившееся направление этой ветке задавал не MetaDriver и не я, а вейсманисты-морганисты (хотя вначале она была узкоспециализированной).
(2) Думаю, что имеет смысл вначале поискать принципы поиска и выбора параметров КК, чтобы в дальнейшем конкретизировать их (наверно, лучше в специально созданной ветке). У меня пока нет никаких идей, как их искать, но надеюсь, что они появятся. Если эти принципы у Вас уже есть в голове, почему бы не обсудить их?
(3) Я считаю, что параметры контекста логично искать в тесной привязке к набору первичных альтернатив ("классу живого существа"). Мне не нравится хаотичный и бессистемный перебор этих параметров в надежде на то, что "когда-нибудь они как-нибудь срастутся". Бесперспективность такого подхода очевидна для меня, когда вижу очередную суперсистему, собранную, скажем, из MACD, Боллинджера, Стохастика и каналов с непонятно как подогнанными параметрами.
1. Well I wouldn't be so resolute about it.... ;)
2. Agreed. About the other branch... not sure. Whatever (almost).
There are a couple of considerations about the principles.
- Phase space coordinates should be more or less orthogonal. (loosely correlated).
For a meaningful search of them we should learn to estimate mutual "orthogonality of ideas" ( OI (c) me), embedded in indicators.
For example, the mashes average and the zigzag opposite highlights the extremes - a suitable pair. There can be many suitable pairs. Let me remind you that the search criterion is a weak (ideally zero) correlation.
- Non-linear indicators have more chances (in terms of profit). Or their combinations with linear ones. (imho, but kind of logical)
// It would be good to write a search engine. More precisely for fast visual estimation of the pairs. The idea has matured, maybe I will scribble it myself. The idea is simple:
// There are three rows on the input - the First indicator (the first phase coordinate (1FK)), the Second indicator (2FK), adjoining future cotier relative to the current
// point (i.e. "correct Buy-Sell" at the point). The output is a flat picture, where points of "correct entry" (2 colours, one "Buy", the other "Sell") are plotted along the first and second coordinates
For now it's enough.
3. Still thinking. Better, it seems to me, not to limit the search. Better to let the autopsy show. :)
1. Но бесконечно философствовать не прагматично. Нужно начинать двигаться.
2. Штурм выдохнется быстрее, если мы не одолеем ни одной высоты, и не получим пускай крохотный, но практический результат...
Машки с 33 немного шокировали меня.
Давайте разберём по косточкам это "домашнее задание". тем более, что данные может получить каждый.
Как я понял, источником разбиение послужила бимодальность распределения идеального отклика...
и она была решительно устранена.
3. Если подобный алгоритм работает на примитивном уровне, почему б его не попробовать применить шире?
1. It is desirable to move in a meaningful way. Any thoughts on where? (I threw mine in).
2. Well, there's a result, isn't there? I was talking about the Sorento just now. And it's obvious to me that all other fitting systems do the same thing in principle and pick out patterns. Including neural nets.
3. That's what I'm saying. This is where it all started. Actually the purpose (mine, and maybe not only mine) of discussion is to systematize searches. And I have almost no doubts that they will be successful, the only question is to what extent.
1) Под эти понятия, пожалуй, попадают только Системы, способные к самостоятельному обучению (не автооптимизация), не использующие шаблонные заготовки типа пересечений машек.
2) Автозменение параметров машек будет автооптимизацией. (выделено синим)
3) Эволюционным развитием исследований моделей-2 не удастся перейти к моделям-3 (выделено жёлтым), так как всё равно придется отказаться от шаблонных понятий модели -2.
4) Либо сразу заниматься моделями-3, либо не заниматься вообще, т.к. к ним прийти "постепенно" невозможно. ИМХО.
1. I don't understand why auto-optimisation doesn't qualify? Any independent change in responses to stimulus+context pairs = learning-3. That's my understanding.
2. it will. Nevertheless, see 1.
3. Perhaps. However, it is said: "Meditate, my friends, meditate. Yes, if you become enlightened, it will NOT happen as a result of meditation. However, if you do not meditate it will never happen. I subscribe to what has been said, actually. It seems to me that learning the properties of learning-2 (especially its limitations) is a strong catalyst for learning-3.
4 Is it a matter of principle at all?
In my opinion, it is just a barefaced emotion, not grounded in anything. It's a kind of maximalism. It's like: "
- Maybe you'll take it in installments? - vindictive Balaganov asked.
Ostap looked at the interlocutor attentively and answered quite seriously:
- "I would take in parts. But I need it all at once."
С СКО в "домашнем здании" не сложилось похоже..
;).
What do you mean?