Design of Social Agents - Foundations on Natural and Artificial Computation - page 195

plays only defectively. If at this stage the partner made at least one cooperative
move, the agents concludes that exploitation was successful and try to exploit
the partner again. If the agent has already tried to exploit after n steps of coop-
eration and this attempt failed, the next attempt to exploit will start after n +1
steps of cooperation. If the agent tried to exploit after N max steps of mutual
cooperation and this attempt failed, it never tries to exploit again. After every
unsuccessful attempt of exploitation agents tries to establish cooperation. More
details about the algorithm are given in the Tab.1. In this table different columns
and rows represent different inputs (moves of the partner) and internal states
of the agents, respectively. The cells of the tables contain outputs (moves of the
agents) and a new state of the agent. The defective and cooperative moves are
denoted as

and , respectively.

⊕

Table 1. Algorithm of the Constructed Social Agent

Internal State Cooperative input Defective input No input

S

,D10 ⊕ ,C0 ⊕ ,C0

D10

,D10 ,D11

C0

⊕ ,C0 ,CF0

D11

,D10 ⊕ ,C1

CF0 ,D10 ,CF0

C1

⊕ ,C1 ,CF1

CF1 ⊕ ,R11 ,CF1

R11 ,D20 ,CF1

D20 ⊕ ,R11 ,D21

D21 ,D20 ⊕ ,C2

C2 ⊕ ,C2 ,CF2

CF2 ⊕ ,CF2 ,CF2

2.4 Evolution of Agents
In our study we apply an evolutionary optimization to find a well performing
agent that can be compared with the constructed one. The evolution starts from
a random agent containing the same number of states (12) as the constructed
one. At every step of the evolution we perform a random mutation of the agent by
changing one element in one of the two matrices that define the agent behavior.
If the mutated agent performs better as the parent agent, we replace the parent
agent by its child and continue the procedure.

3R su s

An interaction between any two agents can be described by the ordered sequence
of moved made by the two agents. If the interaction between the two fixed agents
is long enough, the sequence of moves converge to a periodic pattern (because
of the limited number of the internal states of the agents). As a consequence
the performance of the agents, which is defined as the average score per game,