offline training Reinforcement learning

I was wondering how can we train RL model offline using the RL toolbox?

I have data and I need to use it to train my RL agents. Is this option available in RL toolbox?

thanks

Answer 1

[Xiangyin Wang]

I am recent doing same work, i build a rlagent ,the observations are from the existed data, then generate a action, the action is compared with the actual action from the existed data. I have some problem with the data using, I hope we can have some discussion.

Answer 2

[Kautuk Raj]

Since MATLAB R2023a, you can train a reinforcement learning (RL) model offline using existing data through the Reinforcement Learning Toolbox in MATLAB.

This can be done by utilizing the trainFromData function. It can useful if you have pre-collected datasets and want to apply offline RL techniques. Details about this function can be gathered from this documentation page of the current version (R2024a): https://www.mathworks.com/help/reinforcement-learning/ref/rl.agent.rldqnagent.trainfromdata.html

Further, this example on training a reinforcement learning agent offline to control a Quanser QUBE pendulum can also be looked at to understand the usage of the trainFromData function: https://www.mathworks.com/help/reinforcement-learning/ug/train-agent-offline-to-control-furuta-pendulum.html