Weapon Selection Algorithms Based on Neural Networks for 3D First Person Shooter in Unreal Engine 4

This paper deal with development of artificial intelligence for video games. The main aim of this work is to improve applications of neural networksfor creating of behavior models of characters controlled by computer in video games, more particularly, in first person shooters. The structure of the diploma consists of several parts. In the first part various approaches for development of artificial intelligence for video games are considered and some flaws of methods employed in the majority of modern game projects are enumerated.We also provide necessary theoretical background. The second part includes description of the model that was chosen for implementation and the learning algorithm for this model.In addition, we describe in details modifications to improve learning quality of computer player.Then we consider the problems arising from implementation of algorithm in Unreal Engine 4 and the description of the project that was created using this game engine. Finally, we compare the results of original and modified algorithms and interpret them in terms of average reward values. The conducted analysis confirms that modifications of the original learning algorithm improve its quality greater than 10% in general and greater than 50% for special class of short-range weapons. We also state the suggestion of how one can balance weapon properties using the same learning algorithm.