Call for Papers
Machine Learning Journal
Special Issue on Machine Learning and Games
http://www.aicml.cs.ualberta.ca/_MLJ/cfp.htm
Scope
Authors are invited to submit full papers presenting original results on
any aspect of machine learning and games. An ideal contribution to
this special issue would be strongly motivated by applications to
commercial or classical games and focused on research issues relevant
to the topics described below. Papers specific to game theory,
however, should be submitted to a forthcoming special issue.
Background
Games, whether created for entertainment, simulation, or education,
provide great opportunities for machine learning. The variety of
possible virtual worlds and the subsequent ML-relevant problems posed
for the agents in those worlds is limited only by the imagination.
Furthermore, not only is the games industry large and growing (having
surpassed the movie industry in revenue a few years back), but it is
faced with a tremendous demand for novelty that it struggles to
provide. Against this backdrop, machine learning driven successes
would draw high-profile attention to the field. Surprisingly however,
the more commercial the game to date, the less impact learning has
made. This is quite unlike other great matches between application
and data-driven analytics such as data mining and OLAP.
There is a broad and familiar spectrum of research relevant to games
applications, ranging from inference in partially observable worlds to
representational issues to faster and more robust methods for speech
recognition. There are a few relatively new research thrusts as well.
For example, online learning in which models are constructed and used
on the fly from data unavailable until gameplay time, is a very rich
source of new and interesting problems. Topics of particular
importance for successful game applications include the generation of
new practical and theoretical tools to help with:
* learning to play the game: game worlds provide excellent test beds
for investigating the potential learning has to improve
agents' capabilities. The enviroment can be constructed with
varying characteristics, from deterministic and discrete as in
classical board and card games to indeterministic and continuous as
in action computer games. Learning algorithms for such tasks have
been studied quite thoroughly, but recent improvements are of
interest for this special issue.
* learning about players: opponent modeling, partner modeling,
team modeling, and multiple team modeling are fascinating,
interdependent and largely unsolved challenges.
* model selection and stability: online settings lead to what is
effectively the unsupervised construction of models by supervised
algorithms. Methods for biasing the proposed model space without
significant loss of predictive power are critical not just for
learning efficiency, but interpretive ability and end-user confidence.
* optimizing for adaptivity: building opponents that can just barely
lose in interesting ways is just as important for the game world as
creating world-class opponents. This requires building highly
adaptive models that can substantively personalize to adversaries
or partners with a wide range of competence and rapid shifts in
play style. By introducing a very different set of update and
optimization criteria for learners, a wealth of new research
targets are created.
* model interpretation: "what's my next move" is not the only query
desired of models in a game, but it is certainly the one which gets
the most attention. Creating the illusion of intelligence requires
"painting a picture" of an agent's thinking process. The ability
to describe the current state of a model and the process of
inference in that model from decision to decision enables queries
that provide the foundation for a host of social actions in a game
such as predictions, contracts, counter-factual assertions, advice,
justification, negotiation, and demagoguery. These can have as
much or more influence on outcomes as actual in-game actions.
* performance: resource requirements for update and inference will
always be of great importance. The AI does not get the bulk of the
CPU or memory, and the machines driving the market will always be
underpowered compared to typical desktops at any point in time.
Each submission will be reviewed according to the standards of the
Machine Learning Journal.
Important Dates
Titles and short abstracts due: January 14, 2005
Papers due: February 11, 2005
Author notification: April 1, 2005
Final versions of accepted papers due: June 3, 2005
Publication: Fall 2005
Submission Information
Only electronic submissions will be accepted. Instructions for
submission can be found at http://www.kluweronline.com/issn/0885-6125/
In the text of your electronic submission, please explicitly state
that the paper is for the special issue on Machine Learning and Games.
In addition to submitting the paper to [log in to unmask], please also
submit to the guest editors:
Michael Bowling [log in to unmask]
Johannes Fuernkranz [log in to unmask]
Thore Graepel [log in to unmask]
Ron Musick [log in to unmask]
All inquiries regarding this special issue should be directed to the
guest editors.
|