Computational Intelligence - February 2017 - 8

In a screen capture mode the agent plays via the same
video interface as a human player. The current best
methods use DQNs but many challenges remain in
rapidly learning strategies.

Figure 2 The IEEE CIG 2016 VizDoom competition: agents compete just like human players, choosing their actions by interpreting the video-feed and acting accordingly. http://vizdoom.cs.put.edu.pl/

often have tens of entries, but the challenge we would like to address now, is to
expand this to hundreds or thousands of
entries, with as many entrants.
To achieve this, I propose two
approaches that will help to increase
engagement. The first is to have rapid
entry tutorial-competition combinations at our main conferences. During
the tutorial delegates will learn about
the competition framework, and be provided with clear suggestions on how to
improve on some sample controllers.
They can then spend time honing their
entries, with the results announced and
prizes awarded at the conference dinner.
We have already trialed this with some

success, though the main aspect to work
on is to balance the complexity of
developing an interesting entry with the
time available to do so.
The second approach is to reduce
the amount of coding involved in
entering a competition. Currently,
nearly all our competitions involve
developing software: entrants usually
provide the entry as a software implementation of an AI agent. This makes it
a requirement that entrants are competent programmers. To achieve wider
participation outside of the programming world, we plan to develop an
alternative mode of competition entry,
where an entry is made by interacting

IEEE ComputatIonal IntEllIgEnCE magazInE | FEbruary 2017

with a web application. This may
involve making a series of choices on a
web form, or may involve more complex web interfaces with scripts to
specify neural network architectures,
for example using ConvNetJS: http://
cs.stanford.edu/people/kar pathy/
convnetjs/demo/rldemo.html.
These exercises can follow on from
tutorial material, where the theory of
what might make a strong controller is
presented, and the reader can then apply
this in the form of well-informed decisions (the well-informed entrant should
be able to beat a random guesser!). In
the simplest case, this could be in the
form of a multiple-choice quiz, hence
anyone with a web browser would be
able to read up on an aspect of CI before
making a series of interesting design
choices. They would then observe how
their design leads to success or failure in
some engaging game.
Over the next two years we plan to
try both of these initiatives and monitor
the levels of participation, as well as
pressing ahead with the regular educational activities of the CIS. In summary,
this is the best time yet to be studying
and applying Computational Intelligence to ever more interesting and challenging problems, and the education
committee will be playing its part in
promoting this and in measuring the
effectiveness of our efforts.
References

[1] S. Lucas and G. Kendall, "Evolutionary computation
and games," IEEE Comput. Intell. Mag., vol. 1, no. 1, pp.
10-18, Feb. 2006.
[2] V. Mnih, et al., "Human-level control through deep
reinforcement learning," Nature, vol. 518, no. 7540, pp.
529-533, Feb. 2015.
[3] D. Perez, et al., "The 2014 general video game playing
competition," IEEE Trans. Comput. Intell. AI Games, vol.
8, no. 3, pp. 229-243, Sept. 2016.
[4] D. Silver, et al., "Mastering the game of Go with deep
neural networks and tree search," Nature, vol. 529, pp.
484-489, Jan. 2016.
[5] J. Togelius, "How to run a successful game-based AI
competition," IEEE Trans. Comput. Intell. AI Games, vol.
8, no. 1, pp. 95-100, Mar. 2016.

http://http:// http://cs.stanford.edu/people/karpathy/ http://vizdoom.cs.put.edu.pl/

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