Reinforcement Learning involves understanding how to map various scenarios to appropriate actions in order to maximize a numerical reward. The agent is not directly instructed on which action to perform. It must determine the most rewarding action through a trial-and-error method. Numerous trial-and-error algorithms, along with their corresponding reward architectures, have been proposed in the literature to maximize a numerical reward. However, consider an experimental scenario with different reward architectures, even if the same algorithm is utilized, the results will be dissimilar. Consequently, this motivates us to consider whether only modifying the reward architecture to improve the performance is possible or not. According to the experiments of the karting microgame, we find some interesting properties that will help us improve the algorithm. For example, via some simple physical intuitions to modify the reward architecture, the performance becomes better eventually.