Information Technology Reference
In-Depth Information
the convergence of this algorithm is proven under some assumptions as well. To
validate both the model and the example algorithm, two tests on adaptation
and optimality were carried out, in section 5. Finally, further discussion about
results, model and the example algorithm is done and some future lines of work
are enumerated in section 6.
2
Related Work
2.1 Psychological Preliminaries
In the psychology literature the widely adopted
Kirton Adaptor-Innovator
(KAI)
index [1] proposes that each individual is located in a continuum between “doing
things better” to “doing things differently”; i.e. being “adaptive” or “innovative”.
This means that each person has, in the cognitive level, a degree of susceptibility
or inclination to be innovative in the way they think. A very adaptive person is
one which tries to excel at their job, getting the best result out of the known
situation. On the contrary a very innovative person would achieve a task by
trying different methods, even though some of them would yield dreadful results.
We as persons are located somewhere in between the two ends of this continuum.
Also extracted from psychology -and independently of the KAI index- we
find a model for emotions, called OCC [2]. This model is widely used because
it specifies which actions, events or objects will elicit any of the emotions. It is
a cognitive approach to emotions which embraces twenty two of them, in eleven
different dimensions (all of the emotions are couples of complementary ones, e.g.
hope
and
fear
). The intensity of the emotions is dependent on four variables:
proximity, sense of reality, unexpectedness and arousal. These four variables
would modify the weight of the value for each of the 22 emotions.
2.2 Approximations to Adaptive MAS
Different approaches to Multiagent Learning (MAL) are examined here. Some
of these approximations make use of an explicit coordination mechanism; some
assume certain behavior of the agents -such as that they behave rationally; of
course, every algorithm has its limitations, and its advantages.
There are some algorithms which through an explicit coordination mechanism
deal with the problem of MAL. One of them is Q-Learning with SSA and ABAP
[3] which does not need assumptions on the environment. The core of the algo-
rithm deals with the problem of finding the same Nash equilibrium point out
of the many possible equilibria. It assumes the agents will behave rationally to
achieve this goal. Some other algorithms based on game theory -just to mention
some of the well-known ones- and Nash equilibria are: Nash-Q [4] -which needs
the agents to be rational and has some convergence problems-; Nash-DE algo-
rithm [5] -which converges to a stationary Nash equilibrium in the general case,
with some assumptions, but still needs the player to behave rationally-. There
are some other approximations for the
general sum
games but they all need the
agents to behave rationally.
