[Japanese] End-to-End仮想化と生成AIの統合がもたらすゲーム開発の未来

Abstract

This paper demonstrates generative AI and machine learning hold the potential to significantly transform the process of game development, even without direct application to game AI, by introducing several instances of AI deployed at Cygames, Inc. Traditionally, AI has been harnessed within the realm of video games to control non-player characters (NPCs) and to facilitate in-game graphical content generation. With the advent of generative AI architectures, such as Transformer and GAN, there is an increased academic interest in leveraging these models as comprehensive control mechanisms for entire gaming systems. Thus, this paper provides a detailed exposition on the development and application of an end-to-end game environment simulator, called "preforkd", specifically designed to emulate smartphone games at a large scale. The simulator integrates transformer-based generative models, trained on pre-existing gaming logs, to simulate user interaction with games. A cornerstone of this system lies in the comprehensive use of a functional programming paradigm to drive both container-based virtualizations for encapsulating legacy game systems with minimal alterations and a scripting language for controlling the virtualized game environment. Notably, this system is capable of running 190 game processes concurrently on a single server equipped with 96 virtual CPUs due to the efficient computing resource management by a functional programming paradigm. This unprecedented density of game processes means that just 53 cloud servers are required to simulate the activities of 10,000 concurrent game users, representing a significant leap in scalability. Additionally, the dynamic cloning of game service components is facilitated by integrating the cloud infrastructure's low-layer API, enabling automated, round-the-clock debugging of developing game services. With a successful four-year operational history at Cygames, Inc., this groundbreaking system has been pivotal in the automatic debugging of both existing and in-development game titles. It also proves to be a reliable tool for game parameter adjustments. In addition to the preforkd system, this paper briefly introduces Learned Pseudo Random Number Generator (PRNG) model, which is a WGAN-GP model trained on random numbers generated by the Mersenne Twister PRNG, and the novel virtual pad called Kinetics, which estimates finger movements using the simplest machine learning model.

Info

URL(ja): 人工知能

URL(en): Journal of the Japanese Society for Artificial Intelligence

DOI: 10.11517/jjsai.38.5_637

Citation: .bib format