Reinforcement Learning Toolbox RAM increment

When I am running trainings using the Reinforcement Learning Toolbox, I noticed that the RAM usage increases significantly as the number of trainining episodes increases. Why is this happening?

Answer 1

[Gaurav Garg]

Hi Tech,

The RAM utilization is expected to increase significantly.

This is because there are multiple number of complex mathematical calculation (e.g. matrix multiplications, matrix inverses, activation function calculation, calculation of gradients) needed to train/test any deep neural network.

Having said that, you can run the trainings on a GPU, which would not only not use RAM, but also increase the speed of training (since, GPUs are best fit for such jobs). You can look at an example on how to train RL netowrks on GPU here.

[Tobias Schindler]

How does this explain that RAM usage increases with increasing the number of episodes in an RL setting? The referenced computations, e.g., matrix multiplications, are independent of the number of training episodes as they are done in each time step and fixed w.r.t the NN architecture.

Please elaborate regarding the increasing RAM usage of the RL toolbox with increasing training episodes as this is a common problem and this (unanswered) question is a google result.