Archive Number: 3.184
Status of Thesis: Completed
Date of start: 2010-01-14
Date of submission: 2010-07-13
Evolutionary Robotics is a process where robotic control systems are developed, following the example of natural evolution to achieve a desired behavior. This thesis presents an approach to reduce the loss of desirable behavior during an evolution process through a stabilization mechanism. Already evolved desirable behavior may get lost again due to random changes of the controller, called mutations. A quality criterion for parts of a controller is introduced that indicates their respective contribution to the robot behavior. Mutations are channeled to affect the
parts, that contribute to good behavior, with lower probability than others. As a result, controller parts that contribute to a good performance were stabilized and the evolved desirable behavior is maintained. This leads to improvements in the number of successful evolutions compared to previous studies.
One objective in Evolutionary Robotics is to achieve a sophisticated robot behavior. This can be the case, if a robot solves several tasks at once. In the second part of this thesis the stabilization mechanism is used to combine different controllers, that solve separate, independent
subtasks. In doing so, the quality criterion is used as a stabilization factor for the input controllers to protect them from loosing their original functionality. Evolution is then used to find a suitable connection for the subtask controllers.