Discussion of article "Neural networks made easy (Part 33): Quantile regression in distributed Q-learning"


New article Neural networks made easy (Part 33): Quantile regression in distributed Q-learning has been published:

We continue studying distributed Q-learning. Today we will look at this approach from the other side. We will consider the possibility of using quantile regression to solve price prediction tasks.

A training model was created using the NetCreator tool. The model's architecture is the same as the architecture of the training model from the previous article. I have removed the last SoftMax normalization layer so that the model results area can replicate any results of the reward policy used.

As previously, the model was trained on EURUSD historical data, H1 timeframe. Historical data for the last 2 years was used as a training dataset.

The work of the trained model was tested in the strategy tester. A separate EA QRDQN-learning-test.mq was created for testing purposes. The EA was also created on the basis of similar EAs from previous articles. Its code hasn't changed much. Its full code is provided in the attachment.

In the strategy tester, the model demonstrated the ability to generate profits in a short time period of 2 weeks. More than half of the trades were closed with a profit. The average profit per trade was almost twice as large as the average loss.

Model testing graph

Model testing results

Author: Dmitriy Gizlyk

happy side  
Are the convolutional layers inputs for part 32 the same as part 27 but just removing the last layer to make it 3 layers ?