Information Technology Reference
In-Depth Information
programming) offer interesting opportunities for creating intelligence in strategy
or role-playing games and, on the Internet, it is possible to find a number of
articles related to the use of evolutionary techniques in videogames. For instance,
[8] shows how EAs can be used in games for solving pathfinding problems; also [3]
focused on bot navigation (i.e., exploration and obstacle avoidance) and proposed
the use of evolutionary techniques to evolve control sequences for game agents.
However, till recently most of the work published on the use of EAs in games was
aimed at showing the important role EAs play in Game Theory and, particularly,
their use in solving decision-taking (mainly board) games [9,10,11,12,13,14].
Evolutionary techniques involve considerable computational cost and thus are
rarely used in on-line games. One exception however, published in [15], describes
how to implement a genetic algorithm used on-line in an action game (i.e. a
first/third person shooter). In fact, the most successful proposals for using EAs
in games correspond to off-line applications, that is to say, the EA works on the
user's computer (e.g., to improve the operational rules that guide the opponent's
actions) while the game is not being played and the results (e.g., improvements)
can be used later online (i.e., during the playing of the game). Through oine
evolutionary learning, the quality of opponent intelligence in commercial games
can be improved, and this has been proven to be more effective than opponent-
based scripts [16].
Related with the work described here, [17] applies GP techniques to evolve
bot AI in Unreal TM . Although there are similarities with our second proposal
described here, there are also evident differences: the main distinction is that
bot AI is coded in [17] as finite state machines (FSM) that do not match our
decision trees (even though these FSM were internally coded as trees); also only
2 states were considered and the rules governing the bot AI just admitted two
inputs (translated to our work described here this would represent decision trees
with only two levels of depths). Moreover, a 8-players game was used for the
experimentation whereas we have considered a version of two/three players; in
addition no subjective evaluation was considered and the fitness function was
different to ours. In any case, the results obtained in [17] support our conclusions
that GP is a promising approach to evolve bot AI in FPS games. The same
authors also explored the employment of genetic algorithms to controlling bots
in FPS games [18].
5 Conclusions and Further Work
This paper has dealt with the problem of providing artificial intelligence to
non-player characters (i.e., the bots) in first person shooter games, and more
specifically in the context of the game Unreal Tournament 2004. Two different
proposals based on Decision Trees to code bot AI in UT2004 have been de-
scribed. The first proposal represents the approach that currently is followed in
the development of existing commercial games and consists of manually coding
the bot AI. This way, our manufactured bot AI has been pre-programmed as
a decision tree with multiple rules. The second approach is based on genetic
 
Search WWH ::




Custom Search