update reinforcement policy.m weights
5 views (last 30 days)
Victor Bayer on 15 Jun 2021
in order to run an RLAgent on a Raspberry i have generated a Policy.m file out of the saved Agent (see: https://www.mathworks.com/matlabcentral/answers/854085-run-reinforcement-learning-agent-on-raspberry?s_tid=srchtitle
This file is attached to the question (evaluatePolicy.m).
In the Simulink-model running on the raspberry (Raspberry_USB_.slx) this file is called as replacement to the RLAgent Block, since that one can not be executed on the Raspberry hardware. Through this, an action can be calculated on the raspberry. However, since the Policy.m file does not consider any reward and does not update itself, no learning takes place on the raspberry (see....).
My question is, if there is any way to update the policy function if one considers a reward for the executed action?
The goal is to enable learning on a raspberry.
I am gratefull for any tip.
Thanks and best regards,
Emmanouil Tzorakoleftherakis on 22 Jun 2021
When you want to perform inference on an RL policy, there is no need to consider rewards. The trained policy already knows internally that the actions taken are the right ones.
If you are asking whether you can perform RL training on the raspberry pi, this is not currently supported.