5. RESULTS
5.2. Emergence of Social Behavior
There were 16 1-memory strategies agents in the experiment. In model behavior investigation of IPD, the representative strategies analyzed and discussed could be classified by 4 kinds, ALL-C, ALL-D, TFT, and PAVLOV. Among them, the ALL-C (always cooperate) refers to the ones which would always cooperate with their partners, no matter their partners cooperate with or betray them. On the contrary, ALL-D would always betray and take advantage of their partners. TFT refers to the “repeats the opponent’s previous move” pattern, which cooperates with ALL-C but betrays ALL-D.
TFT is the well-known good strategy in the IPD research. However, its flaw lies in that with the fail to synchronize its memory would cause its despiteful breach against STFT.
The last one, PAVLOV, refers to the “vicar of bray” pattern. The rest of the other 12 strategies could be classified to these 4 patterns. Thus the relations among these well-known strategies are usually used by researchers to discuss complex social dynamics and equilibrium phenomenon. Thus our experiment analysis would take these four patterns as our samples.
z Cellular Automata
Figure 16 is an illustration of the reaction among the 4 well-known strategies in Cellular Automata. In the beginning of evolution, there was no significant difference in quantity. About 3 generations later, we find the quantities of ALL-D agents increased dramatically, while the amount of ALL-C and PAVLOV decreased gradually as a result of being invaded by ALL-D. TFT started to emerge when the quantities of ALL-D reached certain level. TFT would check and balance the growth of ALL-D and coexist with PAVLOV and ALL-C. After approximately 20th generations, TFT would exceed
ALL-D in quantity. Under the pressure of TFT, ALL-D would decrease rapidly. During about the 30th generation, when TFT has grown to certain amount, the memory asynchronous problem of TFT against STFT started to emerge (thus began the vicious circle of despiteful breach). Then PAVLOV would start to increase, because it does not has the problem of failure in memory synchronization. In generation 60, amount of TFT had been less than ALL-D, so ALL-D started growing again. Meanwhile, PAVLOV would decrease as ALL-D increased. At last, in generation 80, TFT would exceed ALL-D again, and the artificial society would reach a dynamic equilibrium, in which amount of PAVLOV and ALL-C were kept stable (evolutionary-stable-strategy, ESS), while ALL-D and TFT checked and balanced each other.
ALL-C TFT
PAVLOV ALL-D
Figure 16: Four well-known strategies in Cellular Automata
Next, we will discuss our experiment group in Cellular Automata. As shown in Figure 17 and Figure 18 are four well-known strategies with mixing ratio 0.1 and 1.0 in Cellular Automata.
ALL-C TFT
PAVLOV
ALL-D
Figure 17: Four well-known strategies in CA ( Mixing self-aware agents with ratio 1.0 )
ALL-C TFT
PAVLOV ALL-D
Figure 18: Four well-known strategies in CA ( Mixing self-aware agents with ratio 0.1 )
According to Figure 17, we find that ALL-D disappeared at the beginning of evolution when Cellular Automata was filled with self-aware agents. Since ALL-D does not match social good expected strategy, our self-aware agent would discover that the existence of ALL-D is not permitted by superego. Thus the self-adjustment mechanism of self-awareness model would start. In order to meet the social expectation, strategy modification would begin. Therefore, in about 3rd or 4th generation, evolutionary equilibrium would be accomplished.
Figure 18 is the main point of this experiment, in which we put self-aware agents into Cellular Automata in the proportion of 0.1. Through our observation, we find that in the beginning of evolution, strategy ALL-D were not as vigorous as what we see in Figure 16 (control group). It has been controlled by self-aware agents. Furthermore, comparing the quantities of high peak (in about 15th generation) of ALL-D in these two figures, we find there were 700 ALL-D in the control group (There are 2500 strategic agents in the simulation world), the group without self-aware agent, while there were only 550 ALL-D in the experiment group, the group mixing self-aware agents with ratio 0.1. There existed an obvious gap of 150 in quantities. This is the key to accelerate the progress in getting rid of social vicious circle. Besides, another special phenomenon is that PAVLOV would exceed ALL-D and TFT in certain period of time and then decrease. That is, PAVLOV could not rival against ALL-D, with the increase of its quantities, it would be more easily invaded by few ALL-D.
z Small-World Network
Figure 19 were the reaction among the 4 well-known strategies in Small-World Network. In the beginning of evolution, the reaction would look like the one in Cellular
Automata. The key difference would not emerge until the 30th generation. ALL-D would thus reach an evolutionary stable, significantly lower than the quantities in Cellular Automata. The Shortcuts in Small-World Network would decrease the world separation, so the reaction among agents would get complicated, and the effectiveness of strategies would get stronger. Thus the evolutionary dynamics would be more vigorous and faster than the one in Cellular Automata. Under Small-World, ALL-D would reach its equilibrium in 30th generation, so the other 3 strategies would reach their evolutionary stable gradually.
ALL-C TFT
PAVLOV ALL-D
Figure 19: Four well-known strategies in Small-World Network
Figure 20 is the main point of Small-World experiment, in which we mixed self-aware agents with ratio 0.1 into the environment. After comparing the quantities of high peak of ALL-D in these two figures (Figure 19 and 20), we find there were 480 ALL-D strategies in control group and only 420 ALL-D in experiment group. There
exists a gap of 60 for ALL-D. The gap is not as large as the one in Cellular Automata, in which there existed 150 for ALL-D; however, it is also helpful for getting rid of social vicious circle earlier.
Figure 20: Four well-known strategies in SW (Mixing self-aware agents with ratio 0.1)
After analyzing the emergent behavior of Cellular Automata and Small-World Network, we conclude that as long as a few self-aware agents, they can accelerate the progress in getting rid of social vicious circle and improved the whole social benefits certainly. This verified our self-awareness model in superego level would be a feasible and effective solution to the conflict between public goods and private interests. It also proved that our Agent Cognition Learning Model is feasible. (The details of all experimental results please see appendix D and E.)