The Future of game A.I. @ Imperial College

Imperial college hosted a two-part seminar on Wednesday entitled “The Future of game A.I.” as part of London Games Week. The aim of the seminar was to bridge the gap between A.I. research within academia and the current and future requirements of A.I. within games. It consisted of two parts: the first was a daytime session for the academics and industry professionals to get together and bash heads, and the second was a public discussion of the days findings and presentations from some of the game industry professionals and academics involved. The three speakers were Peter Molyneux (founder of lionhead, father of all god games), Marc Cavazza of the University of Teeside, and Simon Colton of Imperial college’s Combined Reasoning Group.

(Full writeup after the fold)

Peter Molyneux

Peter Molyneux began by discussing the current state of AI within games. He stated that there is very little AI in the games industry and described the current perception within the industry and amongst consumers as effectively being comprised of four elements:

• Navigation
• Avoidance
• Crude Simulations
• Scripted behaviours

While this has been sufficient for game A.I. for now, the next generation of games will have a whole new set of requirements. Peter gave a short list breaking these areas down:

• Agent A.I. – The need to create convincing characters. Rather than simple state-driven agents based on simple requirements and reacting on events, more engaging, believable characters need to be created.
• Cloning A.I. – The need for player mimicking. Peter gave the example of your Xbox live profile being able to learn from the way you play, being able to realistically replace you online while you take a break.
• Learning A.I. – Needed to adapt to player interactions. Nowadays its too easy to learn the patterns and responses of an agent or world. Learning A.I. would allow the game to surprise the player and maintain the suspension of disbelief as the systems remains invisible.
• Balancing A.I. – A great point here, as he discussed the need for a game’s difficulty to adapt to the player, and the impossible task of balancing a game to appropriate challenge for all players. A self-balancing game would continually adjust based on player performance to prevent those shelving-moments in games while preventing boredom. He suggested that this was the reasoin why games have not yet become mass market.

Peter stated that advances in A.I. would fundamentally change the way that games are designed, and allow the creation of entirely new types of games. Advances will also allow players to have entirely unique experiences as each time you play a given scenario it will evolve differently, and will allow far richer, more realistic worlds to be created as more and more elements react more believably.

Marc Cavazza

Marc focussed on the work of his group at Teeside University – Interactive Storytelling. The goals of Marc’s group are:

• Make games more film Like
• Reconcile story and interaction
• Re-incorporate the aesthetic qualities of linear media

Marc’s talk centred around a prototype story generator, currently being employed to create Sitcoms (or possibly soap operas, this seemed to change as the talk went on). Ultimately, the goal is not necessarily to hand over all authorship to the story generator, but allow an author to create a core story but allow for variations within it.

Marc’s method uses Heirarchical Task Network Planning to generate the stories. Roles for each character are defined, and modelled using plans. The dynamic interaction between these characters contribute to generate multiple situations that were not encoded in the original plans. The success of the story generator is being measured through the number of narrative idioms recognised by the simulation. Marc’s ambitious next step is to extend this to generate real dialogue for each interaction.

Another demo that Marc gave was an example of the world/environment as actor. His demo was an Inspector Closeau-esque, with the world adapting to the events of the player to make misery at every turn. Ie you put you try to wash your hands, and the water sprays all over your face. If you try and make toast with your wet hands, you get electrocuted by the malevolant toasted. Fun stuff, but as Marc stated, the challenge with this is its scalability.

Simon Culton

Simon, assisted by Marc Hull (both academics from Imperial), presented “A.I and games…Do’s and Don’ts”. It was a great insight into the differences between the work being done by AI developers in the games industry and the work being done by academics specialising in A.I. Mentioning Introversion (creators of Darwinia and more recently Defcon), there is obviously a heritage of games at Imperial.

Simon began by discussing the unhealthy A.I. obsessions of both the games industry and academia. The games industry is obsessed with modelling opponents, and academia’s focus on games seems to be obsessed with board games.

There are multiple mis-matches – AI opponents have low RAM, low CPU, and very little time for results to be generated. AI agents, however, require exactly the reverse. There is a mis-match between expectations of what opponents should be like and what they can be within the confines of a game.

Simon quoted MIT’s Rodney Brooks, noting that the development of A.I. is like evolution in reverse: Developing systems to play chess is easy, but developing one to avoid a tiger is exceptionally difficult. He paraphrased Peter Molyneux comments that “We don’t need better opponents, we need more engaging opponents”.

Simon suggested four main areas in which A.I. could be be hugely useful for the games industry:

• Data Mining
• Affective Computing HCI - How is the player feeling?
• Automatic Avatars – Cloning A.I. as mentioned by Peter Molyneux
• AI Tools in game creation – Assisting the designer by learning from their creation methods

In reference to the last point, Simon described one situation where genetic algorithms could be used to generate world entities. A designer could, for example, design a number of different chairs. These could be fed into the system to generate a number of new designs based upon these. Many will be rubbish but the designer could pick the ones that work well, effectively teaching the system what works and what doesn’t. This could potentially solve the issue of huge content demands of the next generation, as well as suggesting designs that the designer may never have come up with.

Simon’s last example was extending this to a task that is hugely time consuming and repetative for a designer – the creation of cities. Rather than place every building and create every section of each building, the designer would create a few buildings and the system would learn and could subsequently take over and create the entire city. The designer would refine and this process would iterate.

Lastly, Simon recommended the following to the audience: AI Bites, and the work of John E Laird.

I’ll try and get some of the Q&A’s written up soon

3 Comments

Leave a Comment

Username (required) :
Email (required) :
Web Site :
Comment :