I want to use a Deep Q-Network for a specific problem. My immediate rewards ($r_t = 0$) are all zeros. But my terminal reward is a large positive value $(r_T=100$). How could I normalize rewards to stabilize the training? I think clipping rewards to be in range $[0,1]$ makes training harder because it just forces most values to be near zero.

  • 1
    $\begingroup$ why do you need to scale the reward here? the problem I would be concerned with is that almost all of your rewards are non-zero. sparse reward environments can be difficult to solve. perhaps you should look at the Hindsight Experience Replay algorithm as a starting point for overcoming sparse rewards. $\endgroup$ Dec 28 '21 at 23:12
  • $\begingroup$ Thank you. I now understand what I should do. I need to Define a bigger state which contains smaller state and gradually decrease it's volume so at the end of training it is equal to smaller state. $\endgroup$
    – bitWise
    Dec 29 '21 at 7:51

Your Answer

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

Browse other questions tagged or ask your own question.