Design of Emergent and Adaptive Virtual Players in a War RTS Game - Foundations on Natural and Artificial Computation - page 379

(i.e., 1 if virtual player wins and 2 otherwise). Fitness function was then defined
as fitness ( x )= 10000 ∗ ( A−B )

C∗D ; higher the fitness value, better the strategy. This
fitness was coded to evolve towards aggressive solutions.

4 Experimental Analysis

The experiments were performed using two algorithms: our initial expert system
(RBP), and the algorithm PMEA (i.e., player modeling + EA) presented in
Algorithm 1. As to the PMEA, the EA uses popsize = 50, p X = . 7, p M = . 01,
and MaxGenerations = 125; mutation is executed as usual at the gene level
by changing an action to any other action randomly chosen. Three different
scenarios where created for experimentation: (1) A map with size 50

×

50 grids,
48 agents in VP army, and 32 soldiers in the human player (HP) team; (2) a
map 54

28,
with 48 VP soldiers, and 53 HP units. Algorithm 1 was executed for a value of
℘ = 20 (i.e., 20 different games were sequentially played), and the RBP was also
executed 20 times; Table 1 shows the results obtained.

× 46, with 43 VP soldiers, and 43 HP units; and (3) a map 50 ×

Table 1. Results: VP win = number of virtual player's victories, HP win =numberof
human player's victories, HP death = average number of deaths in the HP army, VP death
= average number of deaths in the VP army, mov = average number of movements,
and time = average time (minutes) dedicated per game

VP win HP win HP death VP death mov time

map 50 × 50 RBP 4 16 6 7 5345 3.56

PMEA 6 14 7 7 4866 3.20

map 54 × 46 RBP 9 11 4 3 7185 4.79

PMEA 7 13 6 7 5685 3.80

map 50 × 28 RBP 3 17 3 2 6946 4.63

PMEA 6 14 7 6 6056 3.78

Even though in two of the three scenarios PMEA behaves better than RBP,
note that no significant differences are shown; this is however an expected re-
sult as we have considered just one player what means that the player models
obtained in-between two games are likely similar and thus their corresponding
virtual players also are. In any case, this demonstrates that our approach is
feasible as it produces virtual players comparable - and sometimes better - to
specific and specialized pre-programmed scripts.

5 Conclusions

We have described an algorithm to design automatically strategies exhibiting
emergent behaviors that adapt to the user skills in a one-player war real time