The subject of this thesis is the design of a framework, which can efficiently run AI algorithms on Graphic Processing Units. To archive this we use artificial neural networks and parallel programming technics. Although the details of the system are shown with the help of a real implementation, we study problem generally, so we do not concentrate on concrete technologies or programming languages. The main field of application is planned to be in the video game industry, but the framework should be flexible enough to use it in many different kind of simulation.
The first part of the thesis studies the general fields of artificial intelligence and parallel processing. First it studies the history and application (especially in the game industry) of AI research, then then the modern GPUs and their programmability. Finally it examines the parallelization of AI technologies.
In the second part we study the structure and design of the Neural Agent Project neural network computing system. First, we analyze the theory behind the GPU computing, then the structure of the management framework. We examine the topics of design principles, implementation difficulties and ways of application.