How to train RL-DQN agent with varying environment?

The question is related to reinforcement learning based energy management in hybrid electic vehicle (HEV). I am considering DQN-RL for this work. The actions are the control variable for the energy management system which controls the fuel-rate.

In this case, my environment is an HEV with particular driving profile (UDDS). The objective is to train the agent for the energy management system to achieve the final fuel target (desired fuel) at the end of the drivecycle. However, I want to train a single agent for multiple drive profile to achieve the same target in all the cases.

The problem formulation is similar to the paper: Reference paper

I could train the agent for one driving profile, how to train the same agent for multiple profiles?

Note: The reference paper could help to clarify the exact problem. They have trained the agent for 5 driving profile to achieve same desired SOC.

Answer 1

[Emmanouil Tzorakoleftherakis]

What you are describing is actually pretty standard process to create robust policies. To change the driving profiles, you can use the reset function in your MATLAB/Simulink environment definition.

A simple example is here (take a look at the Reset function at the bottom).

[Jillian Eunice Oliveros]

@Emmanouil Tzorakoleftherakis Hello sir. I have a question regarding the Reinforcement Learning toolbox found at this link: https://www.mathworks.com/matlabcentral/answers/1570073-reinforcement-learning-toolbox-how-to-implement-markov-decision-process-mdp-environment-and-dqn. It would be great if you can take a look on it. Thanks!