How to extract the trained actor network from the trained agent in Matlab environment? (Reinforcement Learning Toolbox)

When the agent is successfully trained using DDPG in Matlab environment, if I want to verify the agent, the following codes should be executed according to the tutorial of MathWorks:

simOptions = rlSimulationOptions('MaxSteps',50);
experience = sim(env,agent,simOptions);

Unfortunately, it is not flexible enough for my program. I hope I can extract the trained actor network from the trained agent so that I can obtain the actions by directly inputting the observation vector to the actor network in each sampling step of my robot program for more complex tasks. However, I can’t seem to find the trained actor network from the following variables in the workspace:

Is there a way to extract the trained actor network? If so, how to call the extracted actor network (e.g., what are the I/O formats of the network)?

Answer 1

[Anh Tran]

You can collect the actor (or policy) from the trained agent with getActor. Then, you can use the actor to predict the best action from an observation wtih getAction.

% get actor representation
actor = getActor(agent);
% actor predicts an action given an observation
action = getAction(actor, observation)

If you want to go further, have a look at the Implement REINFORCE from scratch example