The Dailies - Ep. 6 - Multiple Highlights of Multiple Agents
Multi-agent systems are quite en vogue right now, but they have been a topic of interest for games for many years. Games impose a variety of requirements, in terms of designer control, runtime performance, determinism for debugging, to name a few.
In our GDC 2024 demo, we began by showing an empty saloon with one AI character, the Bartender, and then we ran the demo a second time with a fully populated saloon. Walking into that crowded saloon, with characters at different tables, on the balcony, playing piano, and tending bar for a thirsty Sheriff gave a real “wow” moment. The multi-agent part of the demo was about 5 minutes long. We’re sharing a 1 minute glimpse below.
Our western saloon demo exhibits several aspects of multi-agent simulation:
Characters navigate the environment.
Characters can carry props from place to place.
Characters can proactively initiate conversation or interaction with other characters.
The AI Director organizes subgroups of characters into casts, focused on shared activities.
Characters have configurable and dynamic affinities toward other characters.
Characters can hear nearby conversations, and chime in if they have something relevant to add.
Conversations can involve 3+ characters, including human players.
A key question when designing a multi-agent system for games is, should group behaviors emerge entirely organically out of the beliefs and behaviors of individual autonomous characters? Or do group behaviors explicitly exist somewhere in the simulation, and are somehow communicated to organize individual characters into a shared activity?
The early multi-agent examples from many of the LLM-based character solutions skewed toward the everyone-is-autonomous direction. More recently, there are more mentions of AI Directors who make decisions about who can speak next.
The term AI DIrector has been used in games since it was coined to describe the disembodied AI system in Left 4 Dead. The AI Director at that time was focused on controlling the intensity of the game by making decisions about where to spawn zombies and other game objects. The term AI Director has evolved over the years to refer to a system that makes decisions about other aspects of gameplay like quests, story beats, character relationships.
In our case, Bitpart’s AI Director formulates plans (sequences of physical and linguistic actions) that influence the behaviors of dynamic groups of characters that we call “casts.” Academically speaking, these are Joint Plans shared by multiple characters, which may also account for expected behavior of human players. Our approach is influenced by work in games many years ago. Namely the Drama Manager developed for the experimental game Facade, and a GDC 2002 presentation called “Implementing Wittgenstein” from Richard Evans, reflecting on his work on the AI for villagers in Black & White. Related links below.
https://www.gamedeveloper.com/design/gdc-2002-social-activities-implementing-wittgenstein
https://gdcvault.com/play/1022531/Social-Activities-Implementing
If this multi-agent stuff sounds cool / interesting / exciting, and you would like to try Bitpart AI in your games, sign up for our Tech Preview: