Information Technology Reference
In-Depth Information
1000
100000
Classic RL
Social
Classic RL
Social
100
10000
10
1000
1
100
0.1
10
0
50000
100000
150000
200000
0
50000
100000
150000
200000
Simulator Cycle
Simulator Cycle
(a) Failures per agent comparison.
(b) Welfare per agent comparison.
Fig. 4.
System with two significant changes. Results using 32 agents in the system.
notice changes; but also when there are many agents, it is more probable that they
interfere with each other, which would make harder to find a policy in a short
period of time -policies get more complicated as agents are added in the society-.
That explains also the data shown in figure 4(b), where a big difference between
the two types of agents appear. When having more agents in the system it is easier
to have them spread out of the different states, so when the majority agree that
they should behave adaptively, it is more probable that the different states of the
environment have been already explored by some agent inside the organization.
6
Conclusions and Future Work
This paper has presented a new agent model based on three main topics: the
KAI index, emotions and the opinion of the society in which it is immersed.
The communication of the opinion inside the society is a good mechanism to
trigger innovation when a change requires it. The mechanism works better when
the number of agents increases which is very promising in terms of scalability of
MAS using SoWelL as the learning algorithm.
The convergence of SoWelL has been proven, but the agents can converge
towards a non-global optimum. Where this may be a problem for some environ-
ments, it can be alleviated with agents which are even more pessimistic than the
one showed in this paper -being more pessimistic would move the KAI index
towards innovation faster, thus exploring the environment longer-.
Some further work should be carried out to integrate the social welfare mech-
anism in algorithms such as WoLF or one of its variants. Also tests in zero-sum
games would clarify whether this mechanism works well in pure competitive
games. Some scenarios -rock, paper, scissors; matching pennies; etc.- are a good
benchmark in zero-sum games. Related to this line of research is the possibility
of computing the similarity of goals of two agents given their social opinions. In-
tuitively we can say that when two agents have similar opinions, their goals are
similar, with increasing probability as time increases. This similarity would be
semantically equivalent to a composition between trust and reputation between
both agents, and will allow to
distinguish
cooperative and competitive agents
among themselves.
