Decision Tree-Based Algorithms for Implementing Bot AI in UT2004 - Foundations on Natural and Artificial Computation - page 385

NT(Non-terminals) = ClosestWeapon,ClosestArmor, Distance, Closes-
tEnemy, ClosestMedicalKit, ClosestAmmunition, SeeEnemy, VisibleObject,
WeaponLoaded, AmIShooting, WeaponType, AmIUnderAttack, Ammunition-
Amount, HealthAmount.
The meaning of the symbols should be clear from their terms and we will
not describe them in detail. To avoid the well-known problem of bloat in GP a
maximum depth (i.e., 15 levels) was imposed for our decision trees.

3 Experimental Analysis

In this section we show the results obtained experimentally by considering sep-
arately the two proposals described previously and execute a comparative study
in the context of the “2K bot prize” competition. The construction of the game
AI was done for the videogame UT2004 and, although this provides its own
programming language UnrealScript, our programs were coded in Java.

3.1 General Issues

In the experiments we have considered as opponent one of the bots predefined
in Pogamut (i.e., a plug-in for Netbeans that is a free toolkit for development
of bots in Unreal Tournament 2004); more specifically we have considered the
hunter one. This election was based on our experience as players and with the
aim of reaching a more exciting gaming experience. The hunter bot was tested
against each of our proposals separately in a game. This experiment (i.e., this
game) was executed 5 times in 5 different scenarios (i.e., maps). Each game was
executed during 1 minute and a half (note that each game was executed in real
time and this was very time-consuming as unfortunately we couldn't ascertain
how to execute the game off-line).
Two different variants of the fitness function where considered:

rd +50 ∗d
2

(A more aggressive function) f 1 ( x )=( id +50 ∗ frags) −

(A less aggressive function) f 2 ( x )= id +50 ∗ frags
2

− ( rd +50 ∗ d )

Here x represents a candidate solution, 'id' (resp. 'rd') is the damage inflicted
(resp. received) by our bot to (resp. from) the opponent bot (in terms of health
units), 'frags' is the number of enemies annihilated by our bot, and 'd' is the
number of times that our bot was destroyed by the built-in bot. These definitions
for the fitness were obtained according to our intuition and based on our gaming
experience as players of FPS games.

3.2 The Hand Made Strategy

Our first proposal based on the direct encoding of decision trees is referred here
as 'Hand Made bot' (HM). 5 different maps (i.e., TrainingDay, Sulphur, etc and