Non-linear Reward Deep Q Networks for Smooth Action in a Car Game

We formulate non-linear reward functions on deep Q networks in a car racing game by observing the environment (simulator). We aim to control the car movement (action) more smoothly in the game simulator than in the original. Existing studies about deep reinforcement learning maintained either discrete or non-linear reward functions without considering the environment domain, which may lead to illogical car movements. For instance, the car is blocked by three other cars, yet the game still continues by jumping to one of them. To overcome the issues, we define a non-linear reward function to compute the penalty game score based on the distance between the car and the one in front of it. From the game simulator, we surprisingly enjoy the results from the proposed reward function as the car drives more accurately and smoothly than the SOTA models, even at the start of the game point, by showing the smallest number of crashes and no zigzaggy agent movement when the obstacles are far from it.