Improving the Spatial Awareness of Virtual Agents in Games

点击翻译按钮获取中文摘要

Many modern games incorporate artificially intelligent agents, as opponents, companions/followers for the player, or even just as bystanders.  These Non-Player Character (NPC) agents have different roles, but all of them are required to interact with the game environment in a way that makes sense to the player.  Spatially aware NPC behaviours, however, are rare, with most NPCs enacting simple, stereotypical and heavily scripted behaviours that indicate little understanding of spatial properties relevant to their intended activities.  Advanced opponents in games may "take cover" or mimic other strategic behaviours, but they do so based on embedded, manual space designations and very simple (and highly localized) properties of the space. Characters in general are not able to realistically hide, sneak, explore, ambush, or otherwise exploit an arbitrary space, or one with dynamic, changing properties.  NPCs tend to interfere with each other and the player, and make overly simple, apparently irrational choices in contexts not foreseen by the game developer, inducing frustration by contradicting a player's expectation of how an agent should behave. This research aims to improve the technical capability and interactive qualities of artificial agents in modern computer games.  It is aimed at the mechanical benefits, where algorithmic potential exists to better analyze and abstract a virtual space to improve environmental awareness.  "Summarizing" aspects of the world state in different ways to better, and more compactly represent spatial qualities relevant to common movement and strategic planning problems has significant potential for improving the quality of decisions made by an NPC, whether applied to opponents in less-scripted game combat simulation, or to non-combat NPCs that need to align their behaviour with high-level player expectations.  This work addresses challenges in developing abstractions appropriate to different spatial tasks, focussing on core issues related to use of visibility and perception.  This is coupled with efficiency requirements due to real-time computation concerns, typical of interactive environments, as well as the need for practical implementation designs and tooling necessary in rapid development contexts like games. Games have a long history of pushing the boundaries of artificial intelligence.  Modern games continue to raise that bar in presenting virtual contexts that represent agents and environments in increasing detail, and thus with increasing need to demonstrate rational and effective behaviour.  This research work benefits Canada's strong presence in the game development industry, and offers abstractions and techniques useful for the practical use of artificial intelligence in the growing, real-world integration of virtual environments.