Decision Tree-Based Algorithms for Implementing Bot AI in UT2004 - Foundations on Natural and Artificial Computation - page 386

Table 1. Summary with different fitness strategies

Fitness Training Dodging Sulphur Metallurgy Morpheus Average

Day Map

+aggressive
Hunter/Best 0.8205 0.8120 0.1045 0.3130 0.2280 0.4556

Hunter/Worst 0.4590 0.0000 0.0000 0.0000 0.0000 0.0918

HandMadeBot/Best 0.5535 1.0000 0.9280 0.3000 0.4125 0.6388

HandMadeBot/Worst 0.0000 0.2540 0.0000 0.0780 0.0000 0.0664

aggressive
Hunter/Best

−

0.9150 0.7115 0.3285 0.3250 0.4700 0.5500

Hunter/Worst 0.3780 0.2600 0.0025 0.0000 0.1065 0.1494

HandMadeBot/Best 1.0000 0.6405 0.6335 0.8215 0.9075 0.8006

HandMadeBot/Worst 0.4395 0.1370 0.1200 0.2970 0.3545 0.2696

5 different executions on each of them were done) and Table 1 shows the best
and average values obtained by our bot as well as the opponent one; the two
different fitness functions mentioned above were also considered. Observe that
the HM behaves acceptably well and beats the hunter bot in most of the games.
We have also evaluated our hand-made proposal according to the bases of the
“2k bot prize”, that is to say, via a subjective evaluation involving a number of
human judges. To do so, we have asked five different persons (i.e., undergrad-
uate students that wanted to participate in the experience) to evaluate which
of the bots behaved in a more human-like way (even though, as mentioned in
the introduction section, it is not easy to define what this means). In this ex-
periment we considered our HM bot facing other two different bots (i.e., the
built-in 'hunter' and 'simple' bots). The judges were informed about the rules
of the “2k bot prize” competition and were said that at least one human was
playing (i.e., they were misled 2 ). The mission of each judge was to select the bot
with a 'more human-like behavior'. Results are shown in Table 2 in which 'B'
denotes 'hunter' or 'simple bot' indistinctly. We can observe that HM is chosen
by a majority of the judges. The main reason for it might be that our HM bot
exhibits a less mechanical behavior than those of the predefined bots. This is
not surprising as HM is based on four complex decision trees (one for each of
the states mentioned previously) that consider more combinations of perceived
values than the predefined bots.

3.3 The GP Approach

We have also executed our GP based algorithm in the map TrainingDay. 16
runs were done with the following parameters: number of generations: 20, Off-
spring size: 10, crossover probability: 0.8, mutation probability: 0.1, and fitness
function: the less aggressive one. Initial population was randomly generated (a

2 Note that these experiments are not really like in the “2k bot prize” competition
since no human players are on the game.