Discussion of article "Applying Monte Carlo method in reinforcement learning" - page 4

New comment
 
Maxim Dmitrievsky:

as you wish :)

but this is trivial stuff. Importance depends on the variance of the fiche (almost always, except for very simple models). Forest doesn't perform any transformations on the fiches, doesn't multiply or divide by each other, etc., it just scatters their values across nodes, so there are no interactions, only separation

The sequence of spreading (how deep the predictor will be in the tree) matters, and it depends on other predictors, including those that are better at spreading....

 
Aleksey Vyazmikin:

The sequence of spreading (how deep the predictor will be in the tree) matters, and it depends on other predictors, including those that are better spread....

If it's a 3-split tree, I guess, if it's a forest, forget it.

or find confirmation of your thoughts in literature, research of professionals. Otherwise, why do something that the model is not designed for at all, make up some nonsense
 
Maxim Dmitrievsky:

If it's a 3-split tree, I guess, if it's a forest, forget it.

or find confirmation of your thoughts in literature, research of professionals. Otherwise, why do something that the model is not designed for at all, to invent some nonsense

You are better with literature, especially foreign literature. Thoughts that greedy tree construction can be wrong I have heard from different speakers on youtube. I don't have any authoritative studies, I only have the results of my experiments, it is possible that I am misinterpreting the results. In general, then what is the point of this article, if you can input all predictors at once and get an almost perfect model?

 
Aleksey Vyazmikin:

The literature is better dealt with by you, especially foreign literature. I have heard thoughts that greedy tree construction can be wrong from different speakers on youtube. I don't have any authoritative studies, I only have the results of my experiments, it is possible that I am misinterpreting the results. In general, then what is the point of this article, if you can input all predictors at once and get an almost perfect model?

to remove noise from the unnecessary ones, to lighten the model.

but the point of the article is actually different, it's actually in the title.
 
Maxim Dmitrievsky:

to remove the noise from the extra ones, to lighten the model.

but the point of the article is actually different, it's actually in the title.

So you make a random selection of predictors and choose the best model from it, but it is not known how good it is from absalut, I propose to make a more reasonable selection, which, hypothetically, can approach the absolute model to a greater extent. The goals are the same, but the tools are slightly different.

 
Aleksey Vyazmikin:

So you make a random selection of predictors and choose the best model, but you don't know how good it is from absalut, I suggest to make a more reasonable selection, which, hypothetically, may be more close to the absolute model. The goals are the same, but the tools are slightly different.

There is no random selection of predictors.

what do you mean by "a more informed search"?

 
Maxim Dmitrievsky:

There is no random selection of predictors

what is meant by "more reasonable search"?

I realise that I read (perceived) one word wrong in the article and misunderstood it, it says "the Monte Carlo method or random sampling of target variables should be considered separately", and I somehow perceived that it is about predictors, I apologise.

A more valid oversampling is an oversampling based on past results, with reinforcement, so to speak.

 
Aleksey Vyazmikin:

I realise that I read (perceived) one word in the article incorrectly and misunderstood it, it says "the Monte Carlo method or random sampling of target variables should be considered separately", and I somehow perceived that it is about predictors, I apologise.

More valid oversampling is oversampling with past results, with reinforcement, so to speak.

I don't get the point, probably not relevant to the topic of the article.

I think I've already written about importans.

 
MetaQuotes Software Corp.:

New article Applying Monte Carlo method in reinforcement learning has been published:

Author: Maxim Dmitrievsky

Thank you. Is it possible to make the training using the GPU instead of CPU?

 
jaffer wilson:

Thank you. Is it possible to make the training using the GPU instead of CPU?

Yes, if you rewrite all logic (RF include) on open cl kernels :) also random forest has worst gpu feasibility and parallelism 

1234567
New comment