[alife] Neural networks. Evolution. Games. Your comments?

[Sebastian Risi]

Neuroevolution - the evolution of weights and/or topology for neural
networks - is a common and powerful method in evolutionary robotics and
machine learning. In the last decade or so, we have seen a large number of
applications of neuroevolution in games. Evolved neural networks have been
used to play games, model players, generate content and even enable
completely new game genres. To some extent, games seem to be replacing the
small mobile robots ubiquitous in evolutionary robotics and simple
benchmarks used in reinforcement learning research.

Julian Togelius and myself have written a survey on neuroevolution in
games, including a discussion of future research challenges. The main
reason is that there was no survey of neuroevolution in games in existence;
the other reason was that we wanted a tutorial overview to hand out to the
students in our Modern AI for Games course.

The preprint is available here on arXiv: http://arxiv.org/abs/1410.7326
PDF: http://arxiv.org/pdf/1410.7326v1.pdf

Now we would like your input on this. We have not submitted the paper yet,
but we plan to submit to TCIAIG once we're reasonably sure that we have
included all major work on neuroevolution and games and that we are not
misrepresenting any of it.

So please feel free to send us your comments, either as a reply in this
mail thread or directly to the authors (sebr at itu.dk and julian at togelius.com).
All kind of comments are welcome. We are particularly looking for
suggestions for important work we might have overlooked. It is completely
OK to suggest that we cite you, or even that we cite you more than we
already have. But whatever you suggest, please try to explain how the work
would fit into the current structure of the paper, and provide full
bibliographic info (preferable in BIBTEX format).

Please note the scope of the paper, which is neuroevolution being applied
in some way to a game problem. By games, we refer to games that people
commonly play as games (e.g. board games, card games, arcade games, racing
games, strategy games) but not purely abstract games such as prisoner’s
dilemma, robotics tasks or non-game benchmarks for reinforcement learning,
such as pole balancing.

Looking forward to your feedback.

Cheers,

Sebastian & Julian
-- 
Dr. Sebastian Risi
Assistant Professor
IT University of Copenhagen, Room 5D08
Rued Langgaards Vej 7, 2300 Copenhagen, Denmark
email: sebastian.risi at gmail.com, web: www.sebastianrisi.com
mobile: +45-50250355, office: +45-7218-5127