“[Video game] AI is still in the dark ages,” Epic CEO Tim Sweeney told a crowd gathered for Games Beat’s 2017 industry summit.
The video game industry has witness a tremendous amount of growth, thanks to the incredible increase in computation power in terms of visual representations. Using the parallel computation ability of GPUs, powerful 3D engines are able to render 3D objects in the world model, which can be used to render vast virtual worlds more efficiently in a real-time manner and simplifying game design.
The rapid development of state-of-the-art artificial intelligence, i.e. deep learning techniques, in the last few years happened because of one important factor: the development of GPUs, a powerful computational tool for 3D gaming. Ironically, even though we have witnessed a machine master the notoriously complex game of Go, the video gaming industry itself have not benefited much from deep learning.
We could summarize and analyze the following reasons as to why the video game companies rarely apply modern machine learning techniques in their products:
1). State-of-the-art deep learning techniques are mainly focused on supervised learning, e.g. learning (mostly visual) labeled objects. Since the quantity and features of objects in a game are pre-defined, this information can be easily obtained from the world model, i.e the pre-designed game environment. It is then not necessary to apply supervised deep learning in the game.
2). Reinforcement learning research is being extensively conducted right now. In most cases, researchers examine their models in Atari games (e.g. ). Thus, reinforcement learning is a promising technique to be applied to games. Such learning techniques close the loop for the state-action-reward cycle, which is exactly what games need to define for the behavior of characters.
3). Other novel possibilities of using deep learning techniques include the use of neural-based automatic speech recognition (ASR) and natural language processing (NLP). Especially if we need the players to verbally control their characters in future VR/AR games. Also, the popularity of VR/AR games may benefit from the deep RNNs applications as well, where the prediction of temporal data is necessary.
Admittedly, there are a few barriers that are currently preventing the gaming industry from using heuristic algorithms for decision making. A successful industrial product needs to strike a fine balance between functions, speed, robustness and reliability. This applies from the controllers in factories to the Falcon rocket of SpaceX. No video game company will tolerate any error in neural training, or the large amount of time and power it takes to train a network. Therefore, due to the large opportunity cost of building and changing AI systems for games, the enterprising players will probably be reluctant to do so at the moment.
However, Sweeney predicts that video game companies will ultimately utilize advanced AI techniques when the VR/AR games (or so-called “The Metaverse” games) become popular, due to advancements of cameras and displays. Since the complexity of these “the Metaverse” will increase exponentially in both spatial and temporal domains, the developers will find it difficult to model complicated NPC behavior system with conversation trees or to pre-defined behaviors through a finite state machine. We can imagine such “metaverse” games to be a more complicated version of Pokemon Go, but the game’s AI system will be much more complex, such that the monsters and the NPCs will be able to interact with users in a natural way. Such NPC interactions in an AR/VR environment should realistically feel the same as human interaction. That is what video game AI systems will focus on.
Hence, we shouldn’t be surprised to see an AR version of DOOM where multiple players controlled by reinforcement learning published in the near future.
Schmidhuber, Jürgen. “Deep learning in neural networks: An overview.” Neural networks 61 (2015): 85-117.
Mnih, Volodymyr, et al. “Playing atari with deep reinforcement learning.” arXiv preprint arXiv:1312.5602 (2013).
Author: Joni Chung | Localized by Synced Global Team: Hao Wang