Decision Tree-Based Algorithms for Implementing Bot AI in UT2004 - Foundations on Natural and Artificial Computation - page 390

programming, and consists of evolving automatically with no human interven-
tion) a set of candidate solutions represented as decision trees.
We have also conducted an experimental analysis and have compared our two
proposals, firstly from an objective point of view according to two specific fitness
functions (defined from our gaming experience), and second from a subjective
point of view according to the basis of the “2k bot prize” competition. Our two
proposals have both advantages and drawbacks. While it is evident that game
industry is demanding automated processes to automatically generate the so-
called game AI, it is also clear that this is a dicult task as it is to be very
dependant on the evaluation phase (and more specifically on the definition of
the fitness function). The diculty of establishing a good fitness function in FPS
games comes from considering the user satisfaction as the value to optimize and
this represents a 'hard' obstacle as this component is not easy to quantify math-
ematically; there exist however some interesting papers that open interesting
research lines - see for instance [19,20].
In this paper our hand-coded strategy provides better fitness values than the
best evolved strategy generated automatically via a standard genetic program-
ming (GP) algorithm. In addition the GP algorithm is time-consuming as the
evaluation function requires simulation in real time. However, there are other
issues that favor the automated process. For instance, the hand-coding is also
a costly process that requires many hours of coding; in addition this process of
trial, error and debugging is also very expensive measured in human resources.
Surprisingly the evolved strategy was better evaluated than the hand-coded one
when we considered a subjective evaluation close to that proposed in the “2k bot
prize” competition. A more detailed analysis on this issue should be conducted
in the future although we have already pointed out some reason for it.
As further research we plan to optimize our GP algorithm by a simple tuning
of its parameters. Also, alternative (and more complex) definitions for the fitness
functions, that take into account user's satisfaction, surely would improve the
evolution of strategies. In this sense, human-guided interactive evaluation might
be useful and thus interactive optimization will be considered in the future. It
would be also interesting to see how the two approaches work in a larger scale
experiment.
Acknowledgements. This work is supported by project NEMESIS (TIN-2008-
05941) of the Spanish Ministerio de Ciencia e Innovacion, and project TIC-6083
of Junta de Andalucıa.
References
1. Johnson, D., Wiles, J.: Computer games with intelligence. In: FUZZ-IEEE,
pp. 1355-1358 (2001)
2. Millington, I.: Artificial Intelligence for Games. In: Interactive 3D Technology.
Morgan Kaufmann, San Francisco (2006)
3. Buckland, M.: AI Techniques for Game Programming. Premier Press (2002)
4. Bourg, D., Seemann, G.: AI for Game Developers. O'Reilly, Sebastopol (2004)