Analysis of Feed Forward Neural Network Autonomous Driving Agents in Games

Artificial neural networks (ANN) have proven to be a good solution for many problems. ANNs are able to create models which can handle the complexity and dynamic of many problems. The conventional methods used for creating autonomous driving agents in games tend to feed a waypoint to the control system of the agent and make the agent move accordingly. This process will limit the performance of the agents due to the fact that these agents will only be able to drive on tracks with available waypoint. On the other hand, by considering the capabilities of learning systems, it is possible to train agents capable of driving on tracks which do not necessarily have a waypoint to follow. In this paper the process of training an ANN agent will be examined. The training dataset is gathered from a privileged hard-coded agent's controller which follows a waypoint. Additionally in order to enhance the performance of the ANN agent, a function used for measuring Risk values from the race track will be introduced in the pre-processing stage. All of these factors are considered by having the playability of the game in mind, since non-player characters (NPCs) are one of the key features which determine the entertainment quality of a game. Unity as an industry tool which provides the necessary assets that are required for a driving game is used as a testbed for the proposed ANN agents.