As promised, The Saturday Papers are back, in a new format with shorter ‘seasons’ of six articles apiece. More info on the series’ future later – for now let’s get into some new research!
A lot of great storytelling relies on the intricacies and weaknesses of human character – a villain lies to further their own ends, an eyewitness misremembers a crucial detail, a fairytale hero forgets the one thing they were told not to do. Of course, all of these weaknesses are exactly the things software is designed to avoid – computers are reliable, accurate, and always follow orders. It makes for great word processing software, but it doesn’t always make for interesting games – so why don’t we try and model these weird human idiosyncracies and see where it leads?
This week we’re reading Towards Characters Who Observe, Tell, Misremember And Lie by James Ryan, Adam Summerville, Michael Mateas and Noah Wardrip-Fruin. James presented the paper at the EXAG workshop last month in California, and you can catch his talk on YouTube here. The paper describes a framework for AI characters that lets them model how humans use knowledge – and also describes Talk Of The Town, a game the authors are developing in order to show off their cool new tech. We’ll mostly be focusing on what the framework can do, and why it’s so potentially exciting for game designers.
To explain what this framework does, I think it’s easiest to talk about the game it’s being used in, Talk Of The Town (TOTT for short). The game begins with a dramatic dilemma – the town’s wealthy and successful founder has just passed away, and their enormous inheritance appears to have been bequeathed to a secret lover that no-one knew about. TOTT is played by two people – one takes control of a descendant of the town’s founder, and must uncover the lover’s identity so they can seize the inheritance for themselves; the other player takes control of the lover, and must hide their identity, spread misinformation and evade identification for the week leading up to the will being read.
Every NPC in TOTT has a mental model, based on the paper’s AI framework, that lets them represent everything they believe about the town and its inhabitants. NPCs can remember a whole variety of facts, including what characters look like, where they work and live, who owns homes or businesses and where they’re located. It also records some metadata about these beliefs – including what evidence they have for the belief, how strongly they hold the belief, and whether the belief is accurate or not (the NPC can’t know this information of course, but the game does).
The game simulates the inhabitants of the town as they go about their lives – each day they might go to work, visit businesses, or just stay at home. Routines will decide who does what, and as people bump into each other they gossip and share information. This leads to knowledge being updated in their mental models. Sometimes this information is totally correct and only shared between two people – but the system also models people overhearing conversations, intentionally sharing false information (only randomly for now, but motives will play a part in future), and sharing beliefs that they have partly forgotten or misremembered.
Because knowledge is so fluid and important to the game, the simulation in TOTT is crucial for building up what people know about the town in the first place. The paper cites Dwarf Fortress as an inspiration here (describing their game as a ‘Dwarflike’) because it simulates the founding of the town and the entire seventy-year history, including where businesses are founded, births, deaths and marriages. As NPCs move around the simulated town and go through their entire lives, they are constantly observing the world and adding beliefs to their mental models, allowing the game to begin with an organic, unpredictable but relatively filled-out town of NPCs with good knowledge. This leads to nice emergent properties, too – children know less about the town because they haven’t lived there for as long, and the elderly know a lot but are unreliable because their knowledge decays faster due to age.
It might sound like this is about modelling social norms, like Versu, another AI framework. Lying to people and gossiping in bars are certainly very important social practices, but the framework is really about knowledge, not people, and how the flow of knowledge through game spaces can be an exciting system and mechanic. Games like L.A. Noire try to capture this idea of knowledge and belief as a game mechanic by writing it into their dialogue as a kind of puzzle. But like a lot of adventure game logic, it relies on the player and the designer thinking about the world in the same way, and when they don’t it can lead to frustration and a total loss of immersion.
Talk Of The Town turns knowledge transfer into a system, which means the rules are set down and the player can test them, understand them, and then use them without feeling like the game is playing by its own rules. Which isn’t to say this is trying to replace good writing or narrative design – far from it. But what it might do is allow us to create new kinds of experience – games where the player has to manipulate NPC knowledge, or games where replay value is dependent on not knowing what the NPCs know at the start of the game. Like all my favourite game technology, TOTT’s framework will enable us to do lots of new things, rather than simply seeking to replace great work being done by people.
Where To Find More
Check out the paper for a full rundown of the different types of knowledge, how knowledge can spread, and what they hope to achieve with their game. I believe the authors are intending for their AI framework to be extracted and used independently, which might mean that one day you can simply plug it into your next game project and start creating characters who can believe, forget and mislead. And don’t forget, you can watch this paper being presented by James Ryan at EXAG here.
Thanks for reading along with the first Saturday Papers of Season Two! Let me know if you want longer posts, shorter ones, if you want a particular topic covered or any other kind of feedback. And thank you to everyone who said nice things about The Saturday Papers before – your lovely responses were what motivated me to get around to doing this again.