Abstract:
The basic principles of evolution have been applied to the solution of technical problems
in a wide range of domains. To successfully apply evolutionary algorithms to these
increasingly complex problems, we must develop effective techniques. Coevolution refers
to maintaining and evolving individuals for different roles in a common task, either in a
single population or in multiple populations. This thesis work is on cooperative
coevolution of multi-agent systems. A novel model, the CCMAS model has been
developed for evolving multi-agent systems which have added a new dimension in the
research of cooperative coevolution. This model addresses major limitations of traditional
evolutionary algorithms such as the issues of problem decomposition, credit assignment to
the cooperating agents, maintaining interdependencies among the agents, keeping the
. population diversified, considering the cost of communication etc by giving suitable
solutions to these problems.
In order to justify the efficacy and accuracy of the model, we have applied the model in
predator-prey problem which is a very complex task that is yet to be solved. The CCMAS
model has been found to be advantageous over other evolutionary method in solving prey
capture task. The success rates seem to be almost 95% or more. The number of
generations required to evolve the predators are also much less and the less number of
moves are required to catch the prey. Two predator strategies are also tested: predators
without any communication and predators with communication. It is found that
communicating predators perform much better than non-communicating predators. Our
work also performs the sensitivity analysis of different system and environment
parameters and on some of the characteristics of predator-prey problem likely to affect the
performance of the model. Finally the resulting graphs comparing these analyses are
presented.
