0
$\begingroup$

I have an online path guidance system that learns from a set of past experiences(trajectories) to provide a guidance for the user on how to cover the given space in the best way, adopting to whatever choice the user take. I have previously applied a second-order Markov model and acquired a successful rate of 73% on correctly predicting the path followed in the testing set.

Now, I would like to try another approach: Markov decision processes(Reinforcement Learning) to implement this system and use the first-order Markov transition matrix as the states transition probabilities for the model. However, from what I've understood the agent in such applications is a simulation and would follow the action given by the learned policy, but in my case the agent(user) might suddenly decides to jump from a state to another without following the given suggestion(policy). Is reinforcement learning unsuitable for my application?

$\endgroup$
0
$\begingroup$

You could restrict the learning algorithm to update its action values only when following the policy. For instance, SARSA uses the current reward and action values from the previous and current steps. It doesn't need to update at every step. Or maybe the user's decision to not follow the policy could be counted as a negative reward.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.