Design of Social Agents - Foundations on Natural and Artificial Computation - page 198

agent C . As the consequence, the second way to compare performances of agents
cannot be used in the evolutionary optimization.
We used the above described procedure to estimate performances of the con-
structed and evolved agents. The constructed social agent has been compared
with 10 thousands randomly generated agents. In more details, for every pair of
the random and constructed agent we calculated the performances of the both
agents while interacting with each other. The obtained performances are shown
in the Fig.3. The x-axis and y-axis correspond to the scores of the constructed
and random agents, respectively. As we can see in the figure the performance of
the constructed agent was always better than those of the confronted random
partner. This represents the fact that the constructed agents cannot be exploited
by their partners.
Fig. 3. Performances of the constructed and evolved agents compared with the per-
fomrances of the random agents
In the same way we have estimated the performance of the best agents found
by the evolutionary optimization (see Fig.3). As we can see in the figure, the
random agents outperformed the evolved one in many cases. In other words the
evolved agent does not have mechanisms that protect it from the exploitation.
As we have shown before, the evolved agent significantly outperformed the con-
structed one in the pool of the random partners (with the average score per game
equal to about 3.6 and 3.2 for the evolved and the constructed agents, respec-
tively). However, in spite of that fact, we found out that in the pair interactions
with the random agents the average performance of the evolved agent was smaller
than those of the constructed agent (about 2.6 and 3.1, respectively). It can be
explained by the fact that the constructed agent interact with each other better
than the evolved agents. The average performance of the constructed agents in
the pool of the constructed partner is about 3.0 while the evolved agents in the
pool of the evolved partner have performance equal to about 1.7.