[4L1-GS-5g] Agents

9:00 AM - 9:15 AM JST(12:00 AM - 12:15 AM UTC)

[4L1-GS-5g-01] Departing from State Management: Scheduler-based AI Enabling Real-time Multitasking in Open-ended Environments

〇Naoki Hamada¹ (1. KLab Inc.)

Keywords:

game AI,multi-task system,task scheduler

In the field of game AI, paradigms that explicitly model an agent's state-action space are dominant. This applies not only to traditional Finite State Machines and Behavior Trees but also to many Deep Reinforcement Learning methods that have gained significant attention in recent years. These methods are effective for single-goal tasks in environments where the entire set of possible states and actions is pre-determined. However, in environments where these sets are undetermined, explicitly modeling the state-action space becomes difficult. Furthermore, in tasks requiring the simultaneous achievement of multiple goals, agents often succumb to frequent interrupts and replanning, making it difficult to achieve any goal. In this paper, we propose a "Scheduler-based AI" to address these challenges. The proposed method controls agent behavior by applying concepts from job schedulers on computer clusters, such as jobs, resources, priorities, affinity, and preemption. This approach enables real-time multitasking in open-ended environments without the need to explicitly model a state-action space. We introduce an implementation of our approach in the AR game "Nishiura Meme Sanpo", where an agent recognizes real-world events and processes multiple tasks in real-time.