Daniel Kudenko, University of York
Daniel will be providing an overview of approaches to interactive Drama, highlighting and summarizing the state of academic research on the subject, present an overview of selected system, and point out to potential avenues in industry collaboration to help you figure out where to start looking.
I’ve seen this presentation (slightly altered) at the AI Games Network talk – my notes from it are here. Still worthwhile taking some notes, I’d like to see it applied with some of the issues we discussed afterwards after – like non-complete knowledge, realtime actions and the use of animations and non-verbal or indirect actions.
First action games had weak stories – “Aliens invading from space” – binary ending (shot down or you die), in Space Invaders. Wing Commander connects the story missions, pre-written, not much flexibility – some decision points but it all converges. The actual missions had little story.
In parallel you had adventure games – Zork and so on – pre-defined and pre-written. Adventure games moved onto LucasArts games – Monkey Island and so forth – still al pre-written. Indigo Prophecy/Fahrenheit had elaborate narrative, multiple perspectives, pre-written (my note: no bloody gameplay!).
Stories in newer action-adventure game has integrated story and missions, with some bits of choice but all pre-written. All stories so far have been pre-written and it isn’t scalable.
Utopia – where we’d like to go? Agency – the players being able to influence what goes on. Scalability – so it can deal with large worlds as well as both longer and shorter stories. Domain independence – should just exchange the knowledge base to apply generation to other types of stories. Should also be interesting, immersive and replayable.
Not trying to create better stories or be better then a human author, but to create stories dynamically at run time better, faster and cheaper.
Surveyed the current state of drama managers – academic ones since there are no others to look at. There are not many 3d virtual worlds, some with interaction with objects (30%), social interaction was variable – big limitations (Facade being the best example but still dubious quality/difficult to use). Replayability was around the 10′s of stories, some give you infinite amount and a good amount just give you one story. Techniques included planning, mainly, with plot graphs, bayesian networks and drama structures also used in bits or overlapping planning.
Going onto GADIN, developed at York, working with Soap Operas, which is good for simple drama and generating infinite storylines that go on and on. Came up with 5 generic categories of dilemmas for the system. The story is a sequence of dilemmas and story events that lead to them.
Betrayal – get best result at expense of friend (eg: cheating on partner)
Sacrifice – worst result to benefit a friend (admit to crime friend accused of)
Greater good – Best result despite benefit to enemy
Take down – Worse outcome to detriment enemy
Favour – Choose which friend to benefit
GADIN plans using a set of knowledge – actions, stories, characters, dilemmas – in a planner, with a user model to guess what the user will do. The user then interacts with that.
The planner has contingency plans since you can’t accurately predict what a player will do. Example was given of getting drunk to provide a dilemma of who to go out with.
Passes the turning test – ie; users don’t know the difference, 42% guessed the human story wasn’t real on a forum. Surveying users, there were good things on interactivity but people didn’t like the text so improvements are needed.
The NPC’s add to the drama that is performed intentionally, and the drama manager can be made goal-directed to direct the story. Not trying to replace authors, who still are needed to create the game world – it helps simplify it all.