mean that the design process for behaviors of NPCs
in a game have to be easy for designers to test, con-
trol and customize to achieve the desired behaviors.
In order to address this issue, we propose a solution
to conceive and efficiently test collective behaviors,
commonly called strategy in RTS games. In this pa-
per we present a model to design strategies accessible
to the game designers. The proposed model aims to be
used to design behaviors that are adaptable, reusable
and reliable. To address the specific issue of entertain-
ment, a semantic layer is added to the behavior model
to clarify the impact of the behavior on the game ex-
perience. After presenting the related work and its
limits, we will present our behavior model and then
discuss the perspectives we foresee for this work.
2 RELATED WORK
2.1 RTS Research
(Ontanon et al., 2013) has gathered most of the work
related to RTS games AI, separating the work from
Starcraft AI competitions in the CIG and AAAI con-
ferences, and the work done for research purposes.
Both round-robin and single-elimination tourna-
ments proposed during the CIG and AAAI confer-
ences focus on the performance of the competitors.
The AI system of each participant competes against
the others and the only parameter that will deter-
mine the winner is the percentage of winning games.
Moreover, the architectures created for the events are
adapted specifically for Starcraft, broken down into
several parallel and hierarchical modules (Ontanon
et al., 2013). It shows the importance of decomposi-
tion of the decision task, but the specialization of the
modules makes it difficult to reuse them in another
environment. With our model, we aim to provide a
strategy structure that can be used in diverse environ-
ments.
The RTS environment has become widely used as
a testbed in game research because of the multiple
challenges that emerge from it (Buro, 2003). Multiple
IA techniques have been tested for their effectiveness
but often lack usability required for commercial use.
For example (Dereszynski et al., 2011) uses sets of
game logs from Starcraft to produce hidden Markov
models. The resulting behaviors depend entirely on
probability which makes it unpredictable and limits
the game designers control. Case-based planning in
(Ontanon et al., 2007), studied more extensively in
(Palma et al., 2011), also fails to provide the neces-
sary control over obtained behaviors and requires ex-
perts to create example libraries. Furthermore, both
use a learning process which does not integrate well
in the creation process of a video game. Indeed, a
learning process can only be performed properly on
a completed game. If an incomplete version is used,
it is faster to restart the learning process from scratch
than to adapt the previous result to the final version.
Automated planning has also been used in game re-
search, (Churchill and Buro, 2011) uses it to optimize
build order in Starcraft. The computing time and the
vast search spaces in RTS games prevent it from be-
ing used for the entire decision mechanism. We would
like to provide a reusable solution where the game de-
signer has control over the AI produced and can un-
derstand the resulting behaviors.
2.2 Defining Fun
A lot of studies have tried to explain the meaning of
fun and how it can be triggered. The most studied as-
pect of fun is the level of difficulty, which needs to
be challenging, neither too easy, nor too hard, to stay
between anxiety and boredom as defined in the the-
ory of flow (Nakamura and Csikszentmihalyi, 2002).
Most of the studies resulted in classifications of kinds
of fun, (Malone, 1980; Lazzaro, 2004). (Read et al.,
2002) defined 3 dimensions of fun: endurability, en-
gagement and expectations. Other studies focus more
on the player, (Bartle, 1996; Bateman and Boon,
2006). When comparing these classifications, we can
find some similarities but no consensus has been made
on which one is the most accurate and their use in the
creation of behaviors is still unexploited even though
fun is at the heart of game design. In our solution, we
want to allow the designer to create a behavior that
reacts and adapts to the player, so that it provides a
fun and interesting experience.
3 PROPOSITION
The goal of our work is to provide an accessible strat-
egy model in order to simplify the design of com-
plex behaviors. A strategy is defined as the decision-
making process of the allocation of available re-
sources, such as agents or objects, to sub-tasks in the
pursuit of an overall goal. Our model aims to facili-
tate the designing of reusable, reliable and easily ex-
tendible collective behaviors through the description
of strategies. It will therefore fit into the creation pro-
cess of a new video game, during which several mod-
ifications of the game mechanisms are applied and re-
quire the adaptation of every element, including the
AI of the NPC. We want the model to be apprehensi-
ble, to the extent that the reason for the occurrence
TowardsaResource-basedModelofStrategytoHelpDesigningOpponentAIinRTSGames
211