Decision Tree-Based Algorithms for Implementing Bot AI in UT2004 - Foundations on Natural and Artificial Computation - page 387

Table 2. Subjective evaluation according to the “2K bot prize” rules; 'B' denotes
predefined 'hunter' or 'simple' bot indistinctly

Round 1 Round 2 Round 3 Round 4 Round 5

Judge 1 HM HM B HM HM

Judge 2 HM B B HM HM

Judge 3 HM HM HM HM HM

Judge 4 B B HMB B

Judge 5 HM HM B B HM

Total HM 4 3 2 3 4

Total B 1 2 3 2 1

test to avoid non-factible candidates was also used). The reason because we only
consideres one scenario is because this algorithm is time-consuming as the candi-
date evaluation has to be done in real time; note that in each generation 10 new
individuals are produced and this means to dedicate 15 minutes per evaluation
in each generation (i.e., globally this algorithm takes 5 hours). Precisely for this
reason no tuning of the parameters were done.

Table 3. Improvement percentage: 16 runs; GP approach

Run Worst Best % improvement Run Worst Best % Improvement

1 0,0500 0,2335 18,35% 9 0,1242 0,5568 43,26%

2 0,2755 0,3155 4,00% 10 0,2885 0,2885 0,00%

3 0,6375 0,6375 0,00% 11 0,1650 0,1650 0,00%

4 0,0950 0,1570 6,20% 12 0,0820 0,1475 6,55%

5 0,2671 0,7786 51,15% 13 0,0350 0,3805 34,55%

6 0,2065 0,2065 0,00 % 14 0,3970 0,4485 5,15%

7 0,1400 0,5926 45,26% 15 0,0000 0,6790 67,90%

8 0,2345 0,6323 39,78% 16 0,0070 0,4172 41,02%

Table 3 displays the improvement percentage that this algorithm provided in
each of the 16th runs. The improvement percentage is measured as the difference
in fitness values between the best initial random candidate generated in the
initial population and the best final solution. Several considerations can be done
here: (1) the improvement seems to be not appreciable if the initial solution is
acceptably good (as for instance in runs 3, or 6); (2) the best overall solution
obtained has a fitness value of 0.7786 (see run 5) and was evolved from an initial
value of 0.2671 representing an improvement of about 52% (Figure 2 illustrates
this evolution). Note however that this solution is worst (according to its fitness
value) than the best solution obtained by the HM bot (i.e., 1.000); initially this
might indicate a clear superiority of HM with respect to the GP algorithm.
However this does not necessarily mean that HM behaves in a more human-like
way than the evolved bot. The next section tries to shed some light on this issue.