In the area of micro robotics a few problems will probably never completely be solved. Due
to the implied miniature of the subjects, it is inevitable to avoid certain constraints in calcu-
lation performance and memory capacity. Therefore, it is important to thoroughly explore the
possibilities in both hardware and software. On the given hardware setup of the Jasmine 3
robots we shall probe a software solution in several variants.
In the field of swarm robotics there are quite a number of mechanisms in order to control the
robots behavior. In this thesis we will deal with the method of neural network control. Each
of the swarm robots will thus be controlled by a neural network that uses the sensor values as
input and influences the rotation speed of either wheel as an output. In between the neural
network will be altered in several ways. Primarily we will follow the goal of creating a simple
collision avoidance behavior. But in the course of this work we will try and use more refined
techniques, on one hand to improve the performance on the given task, on the other to combine
more simpe tasks or to move on to more complex ones. We will go from feedback networks, so
called Hopfield networks, to structually evolving networks, bidirectional neural networks and so
called Hebbian learning networks.
Lastly, in consensus with the swarm idea, we will test a number of ways to use communication
to improve the single robots behavior. Firstly by supplying an input and an output for inter
robot communciation that is controlled by the neural net. Secondly by fitness based selection,
i.e. the neural nets on the less efficient robots will be overwritten. And thirdly we will make
an attempt on recombination of neural networks, as unpromising as it may seem on the first
glance. |
