Ho to speed up RL training

I'm training an RL model on a dataset with 5min candles from 01-2019 to 06-2022. It tooks a week to train the entire dataset and it's too much. I'm using GPU with Tensorflow and CUDA on a 32GB ram workstation. Someone could help me please and find the way to speed up this training phase? It Is not possible optimize a strategy with these intervals of time reuired

Hello Mattia,



Great to know you are backtesting such a complicated model.



Let me explain to you what I think about your problem.



The use of GPU, compared to CPU, is a great advantage in computer speed. Depending on the hardware, GPU can reduce backtesting time as 10 times less than a CPU. However, you might have come to a place where your hardware isn't able to reach a better optimization time. The machine learning models actually can be very time-consuming, even with GPU using.



Some comments:

  1. Your 32 GB Ram looks decent enough
  2. Improve the GPU in case you can
  3. You can check Nvidia's AI specialized computers (NVIDIA Embedded Systems for Next-Gen Autonomous Machines) to train your models in this kind of computers (Those computers are for developers, but you can profit from them in case you want to pass the learning curve)



    I hope this comments help,



    Thanks and regards, 



    Jose Carlos