Flipping Coins to Estimate Pseudocounts for Exploration in Reinforcement Learning