Shonan Reports - Improvisational Worldbuilding AI
Posted May 25th, 2019
In March I was at a Shonan Meeting, a research retreat organised by the Shonan Institute and Japan’s Nippon Institute of Informatics (NII). Researchers from around the world gathered to discuss new directions for research, begin forming collaborations, and talk about the issues facing our fields. You can read my report of the groups I saw here: Day One (AI Curators & Critics) and Day Two (AI for Playground Games). On the final day I took part in a group titled Narrative and World Generation, organised by TU Delft’s Rafael Bidarra. Here’s what we talked about!
Rafael’s original interest was thinking about how world generators and story generators are often very related in their aims and work, but actually making them work together is often much harder. We talked about a lot of games, especially famous world generators like Dwarf Fortress or Ultima Ratio Regum, and existing research into these kinds of task, like Mirjam Eladhari’s work on the stories that players tell about their games. Eventually we got thinking about the way humans perform this task – telling stories to one another, making worlds up. Children are particularly good at it, easily improvising, changing the rules, breaking constraints and introducing new ones. There’s a lot of exciting ideas captured in human storytelling that don’t get as much attention when it comes to AI.
Lots of work has been done looking at narrative and world generation, and for much of this research other generative challenges naturally work their way in. Co-creative tools for generation - like the Sentient Sketchbook which has a human and an AI system engage in a dialogue about the design of a map - also reflect this way of working. If a writer works with a world generator, a similar kind of relationship is formed, where the AI system must respond to creative actions by another person or system, which might constrain their future moves. In many ways, these engagements are not unlike a game in themselves. Thinking about how humans improvise stories, and how co-creation can often be gamelike in itself, led us to think about the many storytelling games that exist, and what they could teach us.
Two whiteboard pictures from the group. Left: some notes on the features of A Quiet Year. Right: some inspirational games and related work. Click to enlarge.
Our game of A Quiet Year, set in a mushroom forest, growing out of the skeleton of a mysterious giant creature.
This workgroup was broken up across two half-days, so a few of us took the opportunity to play The Quiet Year, a storytelling game about building a world and telling a story in it, done in a group of four people. Players take turns to pick questions, answer them about the world, and drive forward the development of a community. The game has a lot of fascinating mechanics that represent or comment on ideas like conflict resolution, consensus-building, improvisational play, and the nature of fragile communities. There’s so many ideas packed into the game, that we came back the next morning excited to talk about it as a bridge between human and AI narrative-world generation.
Could an AI play The Quiet Year? Not without a lot of modifications, probably. It involves asking open-ended questions (an example prompt: “The community sees a good omen. What is it?”), negotiating with other players (you can ask players their opinion about an action you want to take, and they give you feedback), expressing preferences and intent (players can disagree about plans, act selfishly, and enact their own solutions to the community’s problems). But perhaps we could begin to build towards playing such a game, either by making the problem smaller, or changing the nature of the game itself. We hypothesised a hybrid AI-human approach, similar to Centaur Chess, where the AI only has to take on some of the roles in the game. Or possible a revised version of The Quiet Year, with a small language for expressing actions, and restrictions on how questions can be answered.
Building AI to play a game like The Quiet Year is exciting not just because it’s such an unusual game for AI, but because it’s a general generative/creative task, and all of these skills will build towards better tools for co-creative storytelling or even fully autonomous AI that can compose stories and build worlds to situate them in. So we’re going to push ahead and try to build some small prototypes that explore some of the more unusual parts of The Quiet Year. Hopefully we can make a little inroad here or there, and maybe discover some new and weird directions to push in in the future. You can follow the group’s participants to see how this maybe influences the things we do in the future here: Rafael, Gillian, Kyoong-Joon, Amy, and Mike.
I ended the week visiting teamLab Borderless, an amazing installation of interactive digital artworks quite unlike anything I've seen before, but a story for another blog. Thank you Gillian!
Posted March 25th, 2019.