Discussing the article: "Neural networks made easy (Part 40): Using Go-Explore on large amounts of data"
Good afternoon everyone. Has anyone managed to train this neural network? If yes, how did you do it?
I collected phase 1 data for the same period as the author of the article (4 months). I got a bd file of 1.2 GB approximately (190 000 features). Then I started training phase 2. Phase 2 has a default of 100,000 iterations. I tried running phase 2 several times. I also tried setting 1,000,000 and 10,000,000 iterations. With all these attempts, the error that phase 2 shows fluctuates within 1.6 ... 1.8 and doesn't go down. Or it grows to 0.3 (with other bd files). When you run phase 3 (in the tester) it does not confuse the trade. It just stupidly opens a trade and holds it until the test time is over. I tried to run phase 3 in the tester in optimisation mode. I tried making 200, 500, 1000 passes. It does not affect anything. The only thing is that the Expert Advisor either opens a deal a little earlier or a little later and holds it until the end of the test, because of which it can in rare cases close in a small plus. But it does not close the deal itself, but the tester closes it because the time is up. I also tried to change the #define lr 3.0e-4f parameter in the NeuroNet.mqh file to 1.0e-4f or 2.0e-4f, but that doesn't work either. What am I doing wrong?
Can someone please explain how you train it? If possible in detail.
At what error do you go to phase 3 ?
How many iterations do you do with phase 2?
What do you do if the error in phase 2 does not decrease?
At what number of iterations do you start to change anything? What exactly do you change?
Is it normal that in phase 3 the EA just opens a trade and does not try to trade? Does it make sense to train it with phase 3 in optimisation mode?
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
You agree to website policy and terms of use
Check out the new article: Neural networks made easy (Part 40): Using Go-Explore on large amounts of data.
This article discusses the use of the Go-Explore algorithm over a long training period, since the random action selection strategy may not lead to a profitable pass as training time increases.
As the training period of the Go-Explore algorithm increases, certain difficulties arise. Some of them include:
Curse of dimensionality: As the training period increases, the number of states an agent can visit grows exponentially, making it more difficult to find the optimal strategy.
Environmental change: As the training period increases, changes in the environment may occur that may affect the agent's learning outcomes. This can cause a previously successful strategy to become ineffective or even impossible.
Difficulty in selecting actions: As the training period increases, the agent may need to consider the broader context of the task to make informed decisions. This can complicate the task of choosing the optimal action and require more complex methods for optimizing the algorithm.
Increased training time: As the training period increases, the time required to collect enough data and train the model also increases. This can reduce the efficiency and speed of agent training.
As the training period increases, the problem of increasing the dimension of the state space that needs to be explored may arise. This may lead to the "curse of dimensionality" problem, where the number of possible states grows exponentially with increasing dimensionality. This makes state space exploration difficult and can cause the algorithm to spend too much time exploring irrelevant states.
To check the quality and efficiency of the trained model, we test it on training and test samples. It is important to note that our model was able to make a profit on historical data for the first week of May 2023, which was not included in the training set but directly followed it.
Author: Dmitriy Gizlyk