Computational Intelligence - February 2017 - 52

The application to games of classical
AI techniques, such as finite-statemachines, scripting, case-based models,
and decision trees, has a long history and
the techniques continue to be used to
this day. However, the need to depart
from such techniques and begin to use
more advanced AI developed at the
beginning of the new millennium, as the
demand for more dynamic, challenging,
and interactive game design increased
[114]-[116]. This trend continues to
grow, as the maturation of new technologies, such as virtual reality and immersive games, demands more creative and
real-time behavior from artificial agents.
Traditional AI techniques offer simplicity,
a faster response time, and lower demand
for resources and therefore, are often
preferred over the advanced AI/CI techniques. However, such techniques are
mostly inflexible as they are constrained
with their own pre-programmed knowledge. While LCS share many of these
difficulties with their counterparts, they
have major advantages for game developers, such as their design flexibility;
their ability to be applied to real-time
environments; their ability to learn continuously and adapt in dynamic environments; and the model transparency.
VII. Future Directions

A wide range of games exists in all of
the categories discussed in Section IV in
which LCS-based agent approaches
have not been applied. Comparative
studies between LCS and other CIbased agent approaches in popular
games (e.g., Dove and Hawkes, Snow
Drift, Chess, and Go), are warranted.
Simple game environments, such as
RoboCode, can be used as a useful platform for the evaluation and further
development of LCS, especially in
multi-step environments.
It will also be interesting to explore
other types of LCS for games. Specifically, there are limited applications of
Pittsburgh-style LCS in games. Recent
advances in this discipline [117], [118]
provide important research opportunities to be explored.
In this section, we offer ideas for
future directions related to LCS for

games. Each of these ideas represents an
under-explored research area that has
the potential to advance both LCS and
games research.
A. Architectures for Games

One primary advantage of LCS that has
not been fully explored is that these systems come with an architecture. Such an
architecture can offer a structured way
to decompose and manage complex
problems. For example, a large body of
literature exists on evolutionary games
and evolutionary spatial games, which
are dominated by the application of
conventional GA-based agent approaches. LCS-based agent technologies can
contribute significantly to this field by
providing more flexible and functionrich agent architectures that allow agents
to learn from previous interactions, in
addition to being evolutionary.
B. Symbolic Non-Symbolic Dilemma

For most recreational games, possibly
the most important performance-based
objectives are to deliver believable
actors and to win the game. Believability of character is a challenging issue. It
can be handled in a data-driven manner,
through fast exploration of a large
search space to make wise and smart
decisions. This has created a demand for
methods that are fast, can explore a
large space of possibilities, and make
decisive winning decisions. Non-symbolic methods that are fast, such as neural networks [119], deep networks [120],
and Monte-Carlo tree search [121],
[122], offer alternatives to symbolic
methods and have been shown in the
literature to have different levels of success. However, these approaches lack the
expressive power to explain decisions in
classic reasoning. Another way to
achieve believability is to explore symbolic methods. In the absence of optimized implementations and appropriate
architectures [85], these methods can be
slow in contexts requiring believability.
However, they can offer rational decisions that can be explained by traversing
the rules in the knowledge base.
Implementations of LCS (e.g., XCS),
integrate non-symbolic RL and sym-

IEEE ComputatIonal IntEllIgEnCE magazInE | FEbruary 2017

bolic rules mechanisms. In a similar
fashion, it is plausible to investigate a
hybrid approach that uses LCS to learn
meta-strategies while a neural network
or a Monte-Carlo method is employed
to implement the specifics of these
strategies. The combination of symbolic
and non-symbolic methods in LCS has
had successful outcomes in other areas,
such as data mining [112], [123] and
robotics [27]. Other recent work has
explored the use of genetic network
programming [124] and Boolean networks [125] representations within the
hybrid LCS models.
C. Reasoning in LCS for Games

The symbolic nature of LCS offers an
opportunity to reason in games. This
level of reasoning can be used to explain
actions for users, which could help them
to diagnose their performance, train faster, or conduct post-action reviews.
However, the learning ability of LCS
comes with a disadvantage; that is, many
rules can be generated, leading to large
rule sets and hence low interpretability.
Moreover, some implementations, such
as XCS, generate rules locally, and this
causes the rule set to increase rapidly,
with many overlapping rules.
The above problems have been
addressed in the LCS literature, but
these solutions have not been tried in
the application of LCS to games. Some
of these methods clean the population
in an offline model [126], [127], while
others compress the population in real
time as learning occurs [128].
D. Role of LCS Role in Interactive
Simulation

Interactive simulations have many uses for
education, training, and planning [129].
The value of these simulations increases
when the system is able to diagnose users'
actions and offer corrective actions. In this
scenario, the objective is not to learn the
game or the simulation itself, but to build
a model of the user to explore ways of
improving user performance. Therefore,
interaction can be adaptive in response to
the users' performance.
Non-symbolic methods are not
useful for handling this problem type.

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