In-Depth: Stanford Conference Explores The State Of AI
[The Stanford-held AIIDE Conference is intended to showcase the best in academic and industry artificial intelligence, and Crystal Dynamics' Dan Kline was there to document it for Gamasutra, including the first public discussion on believable AI-driven characters in EA LA's Spielberg collaboration LMNO.]
Late last month, international AI developers and researchers alike flocked to the annual Artificial Intelligence and Interactive Digital Entertainment Conference at Stanford University.
Led by researchers Michael Mateas of Facade
fame and Chris Darken of the Naval Postgradual School, the AIIDE conference focused specifically on bringing AI game programmers together with AI academics to share findings and discuss future collaborations.
Thirty talks were given over three days covering topics ranging from NPC reasoning to designer tools to procedural dialogue creation.
The conference was best represented by Steve Rabin, Nintendo developer and AI Game Programming Wisdom editor, and Borut Pfeifer, Lead AI Programmer on EA LA's Project LMNO
Rabin On The State Of Game AI
Rabin's presentation was a great overview of where game AI has been, where it is going, and why this conference was important.
Citing originators such as Pac-Man, SimCity
, and Tamagotchi, he called out the work of several recent developers as signposts: Nintendogs
with its universal speech recognition, F.E.A.R.
and its enemy planning system, Forza
's use of neural nets to drive cars, and Façade
's interactive story and natural language processing.
This year, games continued to show AI as a core part of the industry's future. Grand Theft Auto IV
used Euphoria to create procedural skeletons and generate procedural crowds, Spore
dynamically skinned and animated many-limbed creatures, and Left 4 Dead
and Far Cry 2
created procedurally-driven experience managers.
Rabin talked about the challenges that future development is facing: costs are rising, risk is greater. CPU power is improving, but we haven't found a strong AI use for it. Middleware and shared AI architecture are still seeking their way. Procedural content is becoming even more pervasive, and will continue to impact both audiences and developers.
But Rabin put forth his own challenge for the future: Despite all this, why is AI still allowed to suck? Because, in his view, sharp AI is just not required for many games, and game designers frequently don’t get what AI can do. That was his challenge for this AIIDE -– to show others the potential, and necessity, of game AI, to find the problems that designers are trying to tackle, and solve them.
Pfeifer On AI In EA's LMNO
Pfeifer went right at one of these problems: believable human characters. Project LMNO
, EA LA's codenamed second collaboration with Steven Spielberg following Boom Blox
, is exploring tech that brings "Spielbergness" to character AI. Pfeifer is using Spielberg's understanding of acting to construct a new AI framework built on clear player-AI communication.
He wants characters to be understandable as well as intelligent -– when players talk about "dumb" AI, they usually mean the AI's motives aren't clear. Pfeifer showed a number of Spielberg's movie clips as examples -– E.T. and Elliott learning to communicate, the first murder in Munich, and the golden idol scene from Raiders of the Lost Ark.
He identified six key human qualities in each of these character clips the AI should express: conflicting motivations, reactions, attention, meaning, emotional simulation, and physicality.
But there were three recurring problems in his approach: control and feedback between the AI and animation layers, the mind-body communication problem, and the complexity explosion of assets. According to Pfeifer, handling these common problems opens the gate to realistic characters.
While his work is still in progress, he did put forward some solutions: communicate AI failure upwards rather then handling it on lower levels; don't hesitate to abstract the surrounding environment and its motions for the general cases, so not every NPC gets details on everything; think of your characters as always animating with adverbs; and define contexts that give characters a larger comprehension of their situation and how to react.
For example, inside a classroom there would be one specific set of social norms if it's full, a different set if it's empty, and wholly unrelated reactions when being shot at.
Pfeifer believes that handling these contexts gives authors the ability to create meanings that can be translated into behaviors and more traditional AI, mentioning in the Q&A that "non-humans" are in some way the subject of the AI in LMNO
Since the AIIDE conference took place, layoffs at Electronic Arts Los Angeles may have affected the development of LMNO
, though EA has denied
that it has been canceled.]
The Best Of The Rest
Other presentations on the program went straight at game development challenges, as well. For example:
- John Reeder looked at using reinforcement learning to model and balance notoriously slippery MMO economies.
- An interesting proposal from John-Paul Kelly explored using planning to auto-craft AI scripts for The Elder Scrolls IV: Oblivion
- David Pizzi showed off his work for Hitman: Blood Money
auto-generating walkthrough storyboards for designers.
- Jacob Schrum presented a novel nueroevolution approach to create bots that handle conflicting objectives.
- Mark J. Nelson proposed an exciting use of formal logic for design tools that could quickly test and balance design mechanics, providing a new sort of design-demonstrable.
- M. Renee Jansen and Alejandro Isaza each demonstrated an interesting new approach for pathfinding.
- Christina R. Strong laid out research towards true procedural dialogue.
- Anne Sullivan, Sherol Chen, and David Thue all had exciting data showing that AI experience management created significantly better games. From their work and others, it seems this new sort of AI will be a competitive requirement in many games in the next five years.
- Invited speaker and composer David Cope shared his insights on proceduralism from the last 30 years, generating beautiful new Rachmaninoff piano concertos in Eliza, another signpost in AI's past.
Conclusion - More Collaboration, Please?
Much of this work will likely find its way into game projects and GDC soon. But there was a larger theme at the conference: AI research needs game designers to guide it.
There is fantastic research going on, and the conference is a big step towards getting it to development, but solving AI problems without considering common game design challenges has little utility.
And yet trying to predict the needs of game design without designers is paralyzing. There were encouraging signs at the conference that the game designer-AI programmer-AI researcher relationship is improving.
The next big step for AIIDE is to truly integrate these designer's perspectives and their game AI problems, and then to bring together AI researchers and developers to solve them.
[Dan Kline is an AI engineer with Crystal Dynamics, and Gamasutra would like to thank him for this write-up. His full notes from AIIDE are available at his personal blog.]