In policy gradients, is it possible to learn the policy if the chain of actions is selected and performed manually/externally (e.g. by myself or by someone else who I have no influence over)?

For example, we have four actions, and I choose in the beginning an action 2, and we end up in a given state, then I choose action 4 and we end up in another state, etc. (the actions can follow some logic or not but the question is general; some of the actions will end up with positive rewards).

Can we learn any meaningful policy network from such a chain of actions?

  • $\begingroup$ Hello. It may be a good idea to explain why you would like to do this, and how exactly do you plan to choose the actions. $\endgroup$
    – nbro
    Oct 13, 2021 at 13:09
  • $\begingroup$ It seems that what you're proposing is related to imitation learning, but maybe I am wrong, as imitation learning is really not reinforcement learning but supervised learning applied to an RL problem. You're suggesting that humans provide/specify the behavioural policy. I don't remember now all the details of policy gradients to answer this question, but, in the case of Q-learning (which is off-policy), you could in principle use any exploratory policy that explores enough the environment, so even a policy decided by a human. $\endgroup$
    – nbro
    Oct 13, 2021 at 15:45
  • $\begingroup$ For policy gradient, the policy (i.e., which actions to take) is fully determined by the parameter $\theta$. Then, gradient ascent can be used to optimize this parameter. However, when you have external interference, the assumption that the policy is determined by $\theta$ is invalid anymore. In this case, I guess you need other methods. $\endgroup$
    – user50121
    Oct 16, 2021 at 1:07


You must log in to answer this question.

Browse other questions tagged .