How do I set initial condition for a trained Reinforcement Learning agent (RL Agent)?

I have a trained Reinforcement Learning (RL) agent that I want to apply as a controller of a tank. Actually it is the same as the MATLAB example "Create Simulink Environment and Train Agent", which controls a Water Tank. I have modified the reward function and some other criterias. Now it is trained and working well, although I cannot change the initial action of the agent. It is causing problems because it initiates with an incorrect action.

In the PID block, for example, I can set the initialization parameter and make the controller initial in steady state, but I cannot do the same for RL agent.

Can anyone help with it?

[Andrea Torcianti]

unfortunately I am not able to answer your question, as matter of fact I have a very similar problem! were you able to figure out a solution? If yes, how?

[Abraao Tanan]

I found a work around. I set the initial conditions always to be the same during the training, and in some instant of the training I apply a disturbance or a change of setpoint. So the agent learns to start always from the same point. I am not sure if it is the best solution, but it worked for me.

[Shah Fahad]

Hello @Abraao Tanan,

Could you please comment on how did you set the initial condition of actions from the code ?

Thank you.

[Abraao Tanan]

Hello @Shah Fahad,

I could not set initial conditions, I just set the training enviroment to start always with the same conditions.

[Shah Fahad]

@Abraao Tanan Thank you for your answer. This approach could work depending on the type of application. For me, I am looking for a more generic way of initializing it. I don't know why python is able to do it but not MATLAB.