Coverage
mediamuseum

AI and Avatars Research & Collaboration U.K. Workshop Report

Andrew Armstrong on February 2, 2009

This report of the AI Games Research Network was written by Andrew Armstrong (blog), who attended a game AI workshop in the U.K. recently.

On Monday 12th of January the National Media Museum in Bradford, U.K. held the AI and Avatars Workshop, organised by Bradford Universities Professor Peter Cowling, and sponsored by the EPSRC Industry/Academia Research Network on AI for Games and Screen Yorkshire’s Game Republic Academy. The day consisted of six 20 to 40 minute talks, with question and discussion time. The day was finished off by a large panel discussion. The evening also hosted an event in Leeds by Game Republic called “AI is dead, long live online gaming!” although I didn’t take notes from this.

I’ve uploaded all the slides I managed to get from the day, so a big thanks to all the speakers for providing these. My notes may contain errors, missing pieces or need the slides to complete the overview of the talk (as well as the nature of notes being brief and concise, meaning I deliberately cut down some of the material). Alex recorded the days talks with audio and video — there is a video available online from the last discussion panel “The Future of Collaboration in Player Representation and Engagement”, with others going to be released freely for AiGameDev.com Insiders. (Free registration required.)

High?level character behavior for Interactive Storytelling and other derived applications

Marc Cavazza

Photo 1: Marc Cavazza

Marc Cavazza, of Teeside University, started the day with a talk about NPCs, design and avatars. The initial joint research project was BARDS between Teeside and Eidos. It was to explore high level NPC behavior from the perspective of Interactive Storytelling. It’s a good case study for discussing the problems with AI.

The project aimed to add planning to support narrative consistency and to explore the potential use in upcoming Eidos games. Difficulty getting the strategy accepted – cannot take chances for a new title, lack of expertise with it, all things heard many times before. This meant will experiment with one title and develop an academic prototype to show it off.

The prototype took the novel classical novel Madam Bovery and put it into an interactive story. Why choose this? Most games are action based, not many based on psychology and emotions. The core action is changes of feeling and emotions in Madam Bovery. Need the “ground truth” for characters interpretations – the drafts of Madam Bovery describe all the feelings the characters are expressing, like having the complete picture from the author himself.

How many people have read the book? (2 people from the audience) not a bad amount actually. It is about boredom, about the female lead character dreaming. Noted as the first feminist book now, but put the author in court – so shocking at the time. The NPC’s are modelled (and can have a player jump in and impersonate them) using planning. How did we get the planning domain defined? Extremely specific emotions based off the drafts by Flaubert.

Can change the progress of the novel – can change the ending but how can you do this without altering the intention of the novel? Used speech based interaction to allow the game to work (not much physical interaction). Emotional speech recognition could be used to solve the issue of the limits of speech processing. Only capture the emotional aspects of a piece of dialogue rather then the entire meaning. If the NPC is expecting a response then any other response may be negatively taken (example being saying “Emma I don’t love you” which set her to more duty, saying she never wants to see you again). Feedback from users thought that the emotional response was correct, or noticed, and that they affected the story.

What about testing it on a real game, like Hitman? Moving to design…because they said what we really would like is…exploring possible solutions to a game level – a solution creator, and for checking if they missed any paths. Generate a storyboard from the given actions the search returns. One problem is the world is static now, so this was emulated while the system was generating.

Time for AI: Emotions, Goals, Turn Taking (and more) for Intelligent Actors

Jonanna Bryson

Photo 2: Joanna Bryson

Joanna Bryson, from the University of Bath, described the use of memory for actors. Why does time matter? You can’t think of everything. Memory, sequencing and pursuit of goals (taking the time to get something done).

For memory, you’re like a notebook and write bits of things down to remember things. Can access things quickly in memory – important things, usually more recent things. Recent things are incomplete information though, so understanding the memories can be hard without more information. To make more believable behavior, you need context sensitive behavior. You have recent events but these fade, while knowledge (facts, expectations) get stored.

Tanguy’s research into it has the same situation gone into with different preconceptions/knowledge, and having different reactions form that. Tension and mood are, over the short amount of time (given 3 good pieces of news then 3 bad pieces of news) are constant. The constants that change are happiness and anger – someone who was angry at the start only smiles on the first piece of really good news, while someone who was happy to start with smiles and becomes more happy for all the good news. Mood is long term, while emotions are short term. Behaviors are altered by these.

Action selection means sequencing when there are constraints to what you can do. Worrying about the situation where you’re trying to get one thing done.
You have to gather together modules to provide the correct parts for perception, action and memory, and do the selection of actions by putting these together in objects (eg; objects in C++ design) and work on iteration of the design to build it up. Working from the most simple implementation first. The simple interactions prioritise attacking, moving, responding to attacks, in a CTF game. The behaviors can get more complex (two actions at once – such as moving and attacking).

In summary, you can’t think of everything and forgetting things you don’t need in the future, but emotions are a form of memory.

Time for AI: Emotions, Goals, Turn Taking (and more) for Intelligent Actor
Joanna Bryson
Download PDF

Approaches to Interactive Narrative Generation

Daniel Kudenko

Photo 3: Daniel Kudenko

Daniel Kudenko, from the University of York presented a talk about narrative in games related to AI Avatars. The history of narrative in games goes from stories written in manuals - weak stories. Moving onto pre-written stories (Farenheit), and more recently the story overlaid on a set of actions you can’t alter (Bioshock). But you can’t write paths for every decision…

What do we want? Some way to help the author along (if not replace them). Agency in the game, feeling he is part of the story and does influence the story. You also want scalability so it works for a fair amount of time, and domain-independence from the non-story elements, interestingness/immersion in the story, and re-playability if the game was played again in a different way.

GADIN – Generation of Adaptive Dilemma-based Narrative. Idea is to put a point of focus to be lead towards with story actions, and when achieved and the climax is reached, the point of focus changes. Example used was soap operas to keep it relatively simple (contrasting with classic literature for instance).

Dilemmas in soap operas – should I do A or B? Five generic categories of dilemmas extracted from the soap operas – the story is a sequence of events just to lead to these dilemmas. Events are controlled by drama engine and by player actions. It is scalable so is potentially infinite.

The categories are betrayal (cheating on a partner), sacrifice (admit to crime friend accused of), greater good (giving something up to enemy in order to save self), take down (accepting punishment for injuring/killing enemy) and favour (giving something to a friend). The architecture has inputs from the knowledge based (characters) story actions and dilemmas. The user model (trying to predict if a dilemma will be interesting) and user interacts both ways with the planner, both inputting and getting outputs from it.

As an example, a club scene with Mary (user) and 2 guys, who should end up both interested in her and her interested in both of them (so the eventual decision will be detrimental to one of them). A live example shows us flirting with someone, drugging them then having an affair – as well as reporting a friends stealing crime, where upon she then says you should stop the affair. The non-interactive generated story was overall positive when questioned users, although could be improved a lot (text only interaction is a problem).

Conclusions are that the problem with the authoring bottleneck is GADIN – which is domain-independent (only deals with the story), among other features which make it advantageous. Problems are it takes time to do the planning, and that limits the game worldwide. A solution to this would be to leave action choice to NPC’s, which is emergent drama. However, may lead to incoherent story (where nothing exciting ever happens even). Want a compromise between the director and the characters choice of actions.

Moving into immersive interfaces, the solution to the text based option is a 3D interface (eg: Second Life). Need to translate the high level actions (eg: flirting) to high level actions, and making it real time but with cutting techniques (such as jumping from location to location).

Integrating it into MMORPG’s and social networking sites with a mixture of players and NPCs would be interesting. The challenge is detection of the social relationships and interactions, and then exploiting those (eg: seeing two close characters and putting a third in there to break them apart).

Approaches to Interactive Narrative Generation
Daniel Kudenko
Download PDF

Discussion: Emotional Interaction with Game Players

Daniel, Joanna, Marc

Photo 4: Daniel, Jaonna and Marc answering questions

This was a panel discussion (basically, a concentration of Q&A for the previous speakers) based on the topic of emotional interaction with game players.

Question: Subjectivity – assuming all characters are together in their likes, while really there are different things people like. Thoughts?
Daniel: Not everyone likes Soap Operas but the sandbox mode allows more ways to do things. The domain is specific to the system – to do James Bond you need to move to another domain.
Marc: Can offer different genres, with some people in each genre liking to bring characters together, and others to take them apart. Comedy has failure too, where funny situations can arise from it. Within that genre that is chosen people interact with that genre, it is not scaled up to a stage for multiple dramas going on at once.
Joanna: It could be very easy to make it flexible. Individual customisation of intelligence, getting to what the player wants. If you start profiling the interesting parts of the story, then you could elaborate the plans more around them and the important characters.

Question: 99% of games are seemingly there are to keep a player there occupied (one action after another). More interesting is character development, plots are setup to elicit an aspect of a characters personality (if the plot was changed the outcome of the book would be the same). Possibly the purpose of a game could be to not just achieve things but to show the character understands different situations.
Joanna: Usually no point competition in it, very touchy feely. Allow some players to tweak the situations.
Daniel: I don’t know any game that does it. My children get a lot of fun out of trying the event five times until successful. Current computer games are not addressing it.

Question: Actors in your title Joanna, we’ve seen characters but with actors they bring something to the role. Are we really there in terms of virtual actors yet?
Joanna: You could see a different actors actions from the different AI procedures seen, even though they had the same face in the research shown. There in very basic ways.
Daniel: Do we really want to get real actors? An ideal computer action is a chameleon of behavior, unlike human actors who bring something to the role.

Question:Adapting a story based on what you profile a player as (including gender etc.) when you gather information, and store it to model the personality.
Daniel: Various characteristics get stored based on the user choice for the dilemmas. It allows the story to pick interesting dilemmas. A selfish player can be given more interesting dilemmas (not ones which are giving you an option of helping yourself or a friend since they’ll only help themselves).
Joanna: Collaborative filtering, like Amazon recommended books, allows this to be looked at - privacy issues though.

Question:The cost at producing the assets for the creation of the expansive narrative rather then a linear world?
Marc: The procedural generation of plot can be used with procedural generation of assets. The designers said they don’t like it, it means there only needs to be one designer instead of four. Loss of control too in the design, rejecting the whole procedural generation because of this (as well as partially the issue of quality – 80% of sitcom plots generated are average or not interesting). Why the game industry is conservative is because it is difficult to predict what someone will like – whether it will sell 500,000 copies or not. (asked about machine learning to combat this) Yes, machine learning, but might not have the right data set (more actions in a comedy might not be more funny).
Joanna: You might be able to get something that gets more and more exciting. Myst and Facebook for instance. Lionhead doesn’t want to move towards it however.

Marc: The ability to communicate using natural language is paramount to NPC interaction. Chicken and egg problem for this.

Someone: In the Interactive Fiction community, Gallataya, single NPC interaction except a conversation with a statue. Borderline game, no points.
Another: Is that a game? What makes a game a game?
Marc: A coffee break question!

Question: I wondered about your Turing test, did you polish up or boiling down to abstract concepts?
Daniel: The second one – the abstract parts (movement and suchlike) can be sometimes not part of the interesting story. It is the narrative skeleton of the soap opera story.

Question: If we boil down the real thing, we mechanising what a complex real soap is – and perhaps before the reduction the model doesn’t fit the soap.
Daniel: Giving the author the skeleton to put the meat on is one way to do it. We however tried to stick to a high level narrative and not to a full story with all the details. Is a limit of the system.

Funding Streams for AI/Games research

Simon Colton

Photo 5: Simon Colton

Simon Colton presented a talk which I only made brief notes on (since the majority of the information is contained in the slides), about funding Game AI related research between companies and universities.

Many funding proposals done (15 in 12 months, 12 accepted, 3 rejected unjustly ;) ). Been working with games companies such as Introversion Software (Small independent), Rebellion Developments (300 strong developer), Emote Games (80 strong, social games).

Examples of Introversion funding were wide ranging in both money and the depth of what the required work was. Most of the time Introversion provided a certain percentage of the amount of money for the project. Projects for DEFCON, Subversion, and others.

Check the slides for much more piratical information (such as who does funding, what was available).

Funding streams for AI/Games research
Simon Colton
Download PDF

Areas for Industry/Academic Research Collaboration in Game AI

Mark Morris

Photo 6: Mark Morris

Mark Morris from Introversion Software presented a talk about his experience with academic collaboration, and the potential for it in the future. Looking at how to make my life better, and this will possibly help any other indie developers out there.

For Uplink, the game picks up whether the player wants to save the internet or destroy the internet (jobs can be good or bad for instance). There are actually other hidden agents that are not shown that work in the background. The question was if it would make a more meaningful experience?

The typical academic way is to fir the problem into something else (this isn’t working, lets plug it into Second Life). Maybe it isn’t best applied to Second Life though? Maybe put it directly into a game, with the actual code? Darwinia hasn’t got any AI academic area specifically, but is fun to watch.

Why might we want to work with academics? To get additional resources and bigger teams, and funding helps pays for things. It also helps build up future employees, and access to the relevant research from the academic side. PR/Marketing is also big, but the main reason is to make the games better. Not harder or more advanced, but more copies sold – so more fun. A bigger (better) experience then the last time. Also needs highly reliable results and technology – 80-90% success is not acceptable for games, need a much higher percent. Any opponent AI’s also need to have a good learning curve.

Introversion works on it’s own design process (which might not be applicable to all companies). An idea spawns concept development which is tight and goes on for a while with lots of iterations, then the game is produced and put into QA before launching. Bigger companies usually go back to just before production rather then doing more concept iterations. This is the area where academics can really help with.

Concept development has a small team (one or two) working with technology to create a workable game concept. Ends in the “first playable”. Very open ended (subversion took 5 years). Pushing the boundaries in this time, accepting new ideas to make it unique (like many indies). Is done with rapid development and undefined core systems and API’s.

The production goes into the larger team to create the final game. Still allows some new ideas to be added but usually it is laid out with deadlines. Companies usually know people will like the previous game with more polygons. Game development is high risk and research is high risk. Sky high risk if you multiple them together. How can you manage the risk? During the concept and development phase, what potential ideas are there to work with the academic community to exploit? Then develop code that is amenable to research. If we can define components of the game that academics can improve then allow the academics to work on it, but if it isn’t completed or it doesn’t work the game still has the rest and can be completed.

All is not lost if it’s not finished the first time around – ports, and other opportunities might arise. Defcon AI project was launched a while after launch, and the game is still being developed (currently for an unannounced mobile platform which isn’t the iPhone). The AI was originally just to train against a bit, was going to be made into more of an opponent. Defcon was a bit a of a mess for academic use, but the API was completed. Looking for more students to look into it, since we’re looking for if there is any ultimate strategy.

Subversion is the second case study. Subversion is at the stage of procedural city generation (a technique needed for generating things in various games), Simon stated previously it’ll involve Espionage. We want to blur the lines between procedural and user generated since users can put their own buildings in the city. A question is how a transport system (eg: Monorail) can be added to the procedural generation – will improve the fun, and is an academic question. If it works, great! If not, no worries.

How can we help each other? The industry is to blame, they need to provide sandboxes to the academics with code, and engineering it so it is ameniable for research. Academia should look for practical areas of research to achieve (also remember 90% is not good enough reliability). Academics should work with live game code and recognise the time pressure of the industry. To get a tangible output, should identify areas of potential collaboration, provide an API, allow academics to produce papers and where appropriate assimilate the research into game development.

Areas for Industry/Academic Research Collaboration in Game AI
Mark Morris
Download ZIP of slides

Conclusion

Overall an interesting workshop which was a lot easier to take notes at then the last one. I am thankful to all the presenters who provided slides since these contain specific information I would have missed copying, and the Q&A’s were delightful insights into the collaboration and the areas of research by academics, and the needs of industry, with a balanced set of speakers to reflect that.

I can only hope the next event is as enjoyable and useful. If you are in the UK and interested in the organisation, joining up for newsletters and information is free, so I’d give it a shot to keep up to date on the events at the very least. If you’re a developer who has had your interest piqued into making applications for research funding, make sure to contact them!

Editor’s note: From the two sessions omitted from the report, the first (panel) is already available as a full video recording in the Insider’s area, and the other will be posted shortly. (Priceless registration is required. :-)

Discussion 1 Comments

William on February 2nd, 2009

Thanks Andrew. Great job!

If you'd like to add a comment or question on this page, simply log-in to the site. You can create an account from the sign-up page if necessary... It takes less than a minute!