Designing conversation-based, narrative-driven games with believable AI is a steep challenge of complex systems. When conversation is a primary gameplay mechanic, players are extra conscious of limitations on the system or of situations where the world’s social fabric doesn’t seem to respond naturally.
But many developers continue to push forward in this arena, evolving the numbers of variables and vectors that help game characters feel like individuals with naturalistic behaviors. Several of these spoke about their detailed AI systems in a series of presentations on socially-engaging AI at the Game Developers’ conference this week.
Prom Week’s strides in this arena have earned it an Independent Games Festival nomination for technical excellence; the game by UC Santa Cruz’s Expressive Intelligence Studio specifically aims for a deeper and more complex social simulation than other dialogue-driven games. Players aim for story goals by interacting with other characters, and are tasked with strategizing within the social space.
In Prom Week, characters are defined by sets of traits like compassion or arrogance, by their word choices and speech patterns, and by their pre-defined relationships with characters in the world. Subjective relationship values – different ways characters can be close, and weight values that modify intentions also play a role in the game’s robust, complex AI system.
“We’re not trying to model reality, necessarily. We wrote these targeting a very specific kind of media experience,” explains creator Michael Treanor of UC Santa Cruz’s Expressive Intelligence Studio, presenting the system in a talk on socially engaging AI at the Game Developers Conference. That particular aim was to create a world influenced by teen media set in high schools.
Character responses are determined through an “accept or reject” rule, where the system calculates the most plausible response based on player choices. Most notably, interactions create permanent state changes in the Prom Week world. The AI system differentiates itself well through indirect effects – the cascading consequences of social exchanges.
Emily Short is a veteran of the language-based game space. She and Cotillion co-creator Richard Evans of Little Text People – the studio recently picked up by Linden Labs – spoke about Cotillion
Described as an “interactive comedy of manners”, Cotillion plays with the prescribed social practices common in Jane Austin’s era. Short wanted to enable players to experiment with both appropriate and inappropriate behavior within the world: “We’ve got a sort of Sims-like freedom, but our design is about episodes that are about 30 minutes long.”
The aim was something that feels less like a game, and more like “a piece of interactive text,” she explains. Its interaction features are real-time: “The AI characters are continuing to act whether you are doing anything or not,” Short explains. “You can sit there and wait for a while to see what happens before you jump on in.”
In a further simulation development challenge, Cotillion is multiplayer – and is completely agnostic about which features are run by players and which are run by NPCs. All characters, whether controlled by players or non-players, have the same affordances. The structure relies on a system of beliefs and questions; if one character expresses a belief, the other character can reply in a standard way, offer a belief on the same topic, ask a question or introduce a new topic.
“But this still doesn’t get us to the level of richness and variability that we really want,” says Short. Thus beliefs are also tagged with emotional significance: for example, expressing a negative belief about someone could lead to them being insulted, or mentioning wealth could lead a character perhaps to envy you, envision you a braggart, or decide you’re a good marriage prospect.
The library of effective content can be detailed through a system of specificity: There are layers to the simple emotional state of being insulted, depending on context. The result: Combining beliefs with possible reactions and layers of specificity depending on context creates a massively complex social universe.
On top of this, Short described how characters in Cotillion attain finely-grained characterizations through lists of traits that can be encoded – for example, a character who is drawn as talkative can be designed to frequently prefer response options within the system that lead to her speaking more, and she can be engineered to favor particular topics when they arise, too. A character that likes to give advice is coded to prefer beliefs that are tagged as “correcting.”
“This gives us an interface between the informational content in a conversation and the emotional content of a conversation,” Short notes. “We can generate standard Jane Austen-esque small talk… but that’s not very narratively compelling or what you want at all times.” So scenes with higher stakes can be generating by creating expectations that characters have for their goals and stakes.
Little Text People’s Richard Evans says the work was inspired by Harvey Sacks’ work on conversation analysis, specifically membership categorization devices and turn-taking. A character can be playing many roles at once: For example, Mr. Darcy is friend, brother and ball guest, and has different traits in each setting – he could be a loyal relative, but a poor participant at the ball because he dislikes dancing.
Say a character makes a sycophantic remark at dinner. The system can evaluate how a character like Mr. Darcy would respond based on the various roles that character plays and how well he plays them. The Sims evaluates friendship level and romantic level; like Prom Week, Cotillion works on three – the nebulous “cool” factor, which suggests basically how well one character’s views and behavior agree with another, without actually indicating anything about the two characters’ relationship.
Turn-taking is a very complex system that, through a set of situational preferences, determines who will speak next; someone who’s addressed directly in a group is liable to speak first, and if someone interrupts – speaks out of social order – there may be situational consequences.
Storybricks’ Stephane Bura has been working to address some of the same design challenges as his fellow panelists, with particular attention to the challenge in scaling complex systems. That these massive behavioral webs are hard to scale can be partially addressed, in his view, by creating intentions – “intention” can reduce the number of rules required to create behavior, as an intention can be associated with multiple attributes and share the same scale.
The panelists had much to offer in terms of progress on interaction systems in games, and the room was full of optimism. Bura noted the teasing Skyrim gets from fans for its AI system, and yet: “How awesome is it that you have characters that understand not only theft, but also the context in which theft can occur?” he reflects.