Machine learning in trading: theory, models, practice and algo-trading - page 2809

New comment
 
Алексей Тарабанов #:

No way.

And training with reinforcements?
 
mytarmailS #:
And reinforcement learning?

The topkstarter wrote a paper on DQN on hubr on R.

It should be understood that reinforcement learning is just a tricky optimisation.

It may work in some cases, it may not.

 
Maxim Dmitrievsky #:

I can't find a numpy analogue for R..

..
What exactly do you need? If speed is required, then data table, if fast analogue of dataframes then tibble, if large amount of data then big matrix.

 
mytarmailS #:
..
What exactly do you need? If it's speed then data table, if it's a fast analogue to dataframes then tibble, if it's a large amount of data then big matrix.

I wanted to alleviate Alexey's suffering ) It's certainly easier from the get-go... but still.

 
Maxim Dmitrievsky #:

The topikstarter wrote an article on DQN on the hub on R

it should be understood that reinforcement learning is just a cleverly designed optimisation

It may work in some cases, it may not.

Well in the context of the memory question...
In RL you can update the action matrix and the state or action evaluation matrix forever, it's like memory with constant updating. But I don't know if I'm right.
 
mytarmailS #:
Well, in the context of the memory question.
In RL you can eternally update the matrix of action matrix and the matrix of evaluation of states or evaluation of actions, that is like a memory with constant updating.... But I don't know if I'm right.

it can adjust the states to new data, but it's all at the level or like Mashka, i.e. with a lag.

It's more important to pick a reward, a target, basically. And it will throw the trades in different directions and at each iteration it will get better and better.

 
Maxim Dmitrievsky #:

I wanted to alleviate Alexey's suffering ) It's certainly easier from the get-go... but still

If Alexey wanted to think and listen, it would take 2 minutes with my script....
 
Maxim Dmitrievsky #:

it can adjust states to new data, but it's all level or Mashka-like, i.e., laggy

It's more important to select the reward, i.e. the target, in essence. And it will throw the trades in different directions by itself and at each iteration it will get better and better

memory is an NS with weights trained, you train it at each step, you move the weights a little bit... not much, that's why there is a lag.

and you can't really transfer that to the terminal.

 
Maxim Dmitrievsky #:

memory is an NS with weights trained, you retrain it at each step, move the weights around a bit... not much, so the lag .

.
A trained network or any trained AMO or even Mashka is memory. Yes, you can accept that analogy....

But when you "retrain" a trained neural network, you change the weights, you forget the past in favour of the present....

It's not retraining, it's retraining at every step, like Mashka in the sliding window, she doesn't remember what's outside the sliding window, even though she was there....
 
mytarmailS #:
.
A trained network or any AMO trained or even Mashka, it's memory... Yes, you can take that analogy....

But when you "retrain" a trained neuron, you change the weights, you forget the past in favour of the current....

It's not retraining, it's retraining at every step, the same Mashka in the sliding window, she doesn't remember what's outside the sliding window, even though she was there....

there's a whole zoo of approaches, you can find implementations on github, I saw one for python.

1...280228032804280528062807280828092810281128122813281428152816...3523
New comment
Reason: