Knowledge Discovery and Data Mining in Medicine - Advanced Techniques in Knowledge Discovery and Data Mining - page 188

To solve a complex task requiring teamwork, sharing of roles among agents
is needed. Generally, it is shown that the performance of heterogeneous agents is
better than that of homogeneous agents, because each agent in a heterogeneous
model is specialized according to its role.
This multiagent approach is effective in solving problems that seem to be
unrelated to the concept of agents. Soule applied a multiagent approach to
even-parity problems and linear regression problems and showed that the
performance of this approach is better than that of conventional solutions [13],
[14]. In these experiments, a solution is obtained by collecting each agent's
output.
In this research, we apply such an idea and search for various solutions by
heterogeneous agents, to the domain of knowledge acquisition from data. We
handle the data containing multiple rules and aim to do both clustering of the data
and rule extraction from each cluster. We use a multiagent approach to solve this
problem. That is, the data are divided among agents. This corresponds to
clustering of data. And each agent generates an approximate function for the
assigned data. This corresponds to rule extraction in each cluster. As a result, all
rules are extracted by multiagent cooperation.
To use this approach, however, we need to know the number of rules hidden
in data and how to allot data to each agent. Moreover, as we prepare abundant
agents, the number of trees in an individual increases. Therefore, search
performance becomes poor. The proposed method using ADG can address these
problems.
7.3.2 Automatically Defined Groups
In ADG, agents can take different programs, which are needed to solve the
problem. Moreover, by grouping multiple agents that have similar roles, we can
monitor the increase in the search space. However, we do not know the optimal
number of groups of grouping of agents beforehand, ADG can address these
issues by evolution. The algorithms for ADG are as follows.
A team that consists of all agents is regarded as one GP individual. One GP
individual maintains multiple trees, each of which functions as a specialized
program for a distinct group. We define a group as the set of agents referring to
the same tree for the determination of their actions. All agents belonging to the
same group use the same program.
Generating an initial population, agents in each GP individual are divided
into random groups. Basically, crossover operations are restricted to
corresponding tree pairs. For example, a tree referred to by an agent 1 in a team
breeds with a tree referred to by an agent 1 in another team. However, we consider
the sets of agents that refer to the trees used for the crossover. The group structure
is optimized by dividing or unifying the groups according to the relation of the
sets. Individuals search solutions as their group structure approaches the optimal
one gradually.
The concrete processes are as follows: We assign an agent arbitrarily to two
parental individuals. A tree referred to by the agent in each individual is used for