Are there methods (possibly logical or (how they are called in the literature) relational) that allows for the developmental systems to understand or explain the value of the received reward during the developmental process. E.g. if the system (agent) can understand that reward is by the chance, then it should be processed quite differently than the reward that is just initial downpayment for the expected series of rewards. Credit assignment is the one method (especially for delayed rewards), but maybe there are different methods as well?

Relational reinforcement learning allows to learn symbolic transition and reward functions and emerging understanding of the reward by the agent can greatly facilitate the inner consciousness of the agent and the search process for the best transition and reward functions (symbolic search space can be enormous).

  • 1
    $\begingroup$ Please define or give at least an example of "trying to explain". Meta-explanation is still a research topic. J.Pitrat's blog could give some interesting insights about it. $\endgroup$ Aug 20, 2019 at 23:39
  • $\begingroup$ By "trying to explain" I mean that the system (agent) have some kind of function (reward value, current state of environment) => ((state, action)=>(reward), whether and how to update the transition, reward function), i.e. function that observer the state and reward value and tries to explicitly model the environment and modeling is done by specifying and updating function. $\endgroup$
    – TomR
    Aug 21, 2019 at 7:17
  • $\begingroup$ That should go into your question* $\endgroup$ Aug 21, 2019 at 8:07


You must log in to answer this question.

Browse other questions tagged .