TradeFlow: What's coming up?

5raQ...PSvC

28 Oct 2024

This is the second post on the topic of RL development, on the topic of predicting prices for cryptocurrencies, on using the CCXT library, as well as on open source collaborations and collaboration. The first post was here, where you can understand more about what I'm talking about in general.

What is reinforcement learning? How does reinforcement learning work?

Here, briefly, I continue to understand Reinforcement Learning and suggest that you connect to this project. Let's look at my case with the prediction of actions in the trading environment depending on the last coin price, based on the deep learning technique that I tested and implemented in the project.

Model study

To study it, I took an already developed case of the DL Model, or rather, several examples. I studied the architecture of. Changed the data. I chose the best model that was in the performance. Since there were several best architectures, then there was something to experiment with. In the free and open source repository on GitHub (I will share it), there was a good model that I needed. Now I have found something else interesting and I want to tie the new selected model to the new version of the project, as well as make improvements and implement best practices that I wanted to prepare last year.

What I wanted to do:

use more than one coin, but make learning and prediction more universal
add an agent for trading by making decisions based on context using large language models (I think the o1 model will do a good job with this task)
add work with the Rust programming language - I think this language is the most promising for production systems now, for design and MVP, I of course choose Python, apparently Python vs Rust will
be able to implement a subscription to signals on via the Blockchain, so that it can be passed to people for testing on their coins, thereby selling the results of my intelligent work
to deploy on a scaled infrastructure to ensure trouble-free operation and simple scaling and secure storage - I chose Kubernates.

In the process of working with the source code, analysis and optimization were carried out, which allowed us to abandon outdated practices and switch to the use of modern and effective development methods. Containerization was implemented, which allowed for more flexible and secure deployment of applications.
The training and retraining pipelines were not implemented on the servers, which allowed us to ensure continuous development and improvement of the development model. A demo dashboard for signal processing was also developed, as well as a user terminal for monitoring trading.

What are the assumptions? What would you like to fix? How do I see the development of this project? To be honest, there are a lot of things, there are assumptions related to writing code, scaling, choosing the model architecture and integrating the model. Everything is not clean, some places are not logical at all, and the desire to correct is expensive. Therefore, it is not yet possible to implement or correct this project in the future. I would like to keep it in mind as an intermediate step to the next victories and challenges that will definitely appear on my way.!

So crunch dao, MID+ONE competition, prize pool of 10,000 USDC. Can I apply this knowledge? Of course! how can I implement my model in crunch dao competitions in order to win? I would like to add that DQN is essentially not the best thing in Deep RL, but only the beginning of.

Thanks to colleagues who developed models that I was able to freely reuse in a non-commercial project, thanks to colleagues who made the University of Artificial Intelligence, thanks to Terra AI, thanks to my examiners, managers and opponents. I also want to express my respect to the employees of Senergy University who were able to lead me by the hand to solutions that I can scale. Thank you to my family, parents and daughter! You are the best thing I have!