Computational Intelligence - February 2017 - 7

communities and the general public.
They also make an ideal test-bed for AI
(Lucas and Kendall, 2006) and provide a
range of scenarios, from the ultra-simple
setting of Iterated Prisoners Dilemma (IPD)
to StarCraft (a Real Time Strategy game
with great skill-depth). Note that I distinguish between the complexity of the scenario and the complexity of the game -
the distinction is important. Go is a
simple game in terms of the scenario
(placing stones on a grid in accordance
with a few rules), but has enormously
rich and complex gameplay, and has been
the focus of research for decades. Google
DeepMind's AlphaGo (Silver et al., 2016)
finally achieved performance at least as
good as the very best human players in
2016, and very probably better.
Interestingly, the best StarCraft AI
agents are vastly inferior to the best
human players, and this is one of the next
challenges for Game AI. IEEE CIG and
AAAI AIIDE have run StarCraft AI competitions for many years, and although
progress is being made (each year the
competition winners generally outperform their predecessors) there is still a
long way to go in reaching and exceeding
the ability of the best human players.
Togelius (2016) gives a summary of
recent Game AI competitions, and provides a set of guidelines for running
them: reading I would recommend to
anyone planning to run one. Some of
the most important points include to
publicise the competition well, provide
clear instructions for entrants, and make
the bar for entry very low. The latter
should be achieved by providing some
sample controllers, and some suggestions
for how to improve on the samples.
Beyond this it is also good practice to
provide a continuously running trial
league, and provide a website where
readers can view past runs. These are all
important operational points, but what
about the choice of game? Here we
need to consider what will be of interest
to the research community, what we
wish to learn from the results, and how
to inspire many participants to enter.
A competition series and software
framework that addresses this in a very
general way, is the General Video

Games have always provided a great challenge for AI,
and General Video Game AI takes this to a new level.
Now the agent needs to be able to play a range of
previously unseen games and rapidly adapt.

Figure 1 A GVGAI agent playing a game inspired by arcade classic Sheriff Nintendo. The
agent was written with no knowledge that this particular game existed.

Game AI (GVGAI) system (Perez et al.
(2016), http://gvgai.net). This enables
participants to compete in tens or
even hundreds of games, either 1-player or 2-player and with a variety of
interfaces (full object model or screen
capture) and modes (planning track,
where agents have a forward model to
run fast "what-if " simulations, and
learning track, where agents must
learn what to do without the aid of a
forward model). The outstanding challenge this represents is that entrants
must submit agents for the competitions without even knowing the set of
games they will be playing!
The GVGAI competition series has
great longevity and depth: the interfaces
and framework will remain compatible
for a long time, while the games can
grow in complexity and vary the nature
of the strategic or reactive challenge
required to compete at the highest level.
As the set of agents grows, we can build
a comprehensive map of how different
types of agents compare with each other
across a wide range of games, providing
a highly infor mative educational
resource. The 2-player track is especially
interesting, as it enables a battle of wits

between AI agents on unseen video
games. This is more informative than
comparing the performance of AI
agents with that of human players on
single-player games. For example, the
highly impressive work of Mnih et al.
(2015), involved the construction of a
deep learning system able to learn to
play a range of video games. This
deserves to be tested in a more demanding way, since machines have superhuman reflexes and so may appear
smarter than they really are when using
human-attained high scores as the basis
of the comparison. The GVGAI 2-player
track provides the perfect environment
to test the intelligence of AI agents in a
scientific way.
The model of using screen capture
to enable software agents to play
against each other was used with great
success in the VizDoom competition at
IEEE CIG 2016 - and all leading
entries were based on deep learning
systems. The screen-capture GVGAI
track will take this challenge to the
next level, where the agents will not
know the games in advance.
The competitions we run in conjunction with IEEE CIS conferences

FEbruary 2017 | IEEE ComputatIonal IntEllIgEnCE magazInE

http://www.gvgai.net

Table of Contents for the Digital Edition of Computational Intelligence - February 2017