I am trying to develop a deep q learning based model for playing a game like flappy Bird. Initially, we let the agent do actions at random , and then it learns how to play by getting rewards and penalties. For flappy Bird, the choice of action is either jump or not. And being random at nature the probability will be 50% for each action. But giving jump action in such a frequent manner will always cause it to go too high . (Like tapping the mobile screen, after every alternate or 3rd frame.) And it will almost never will be able to cross the gap, and will never experience a positive reward.