Diplomarbeit DIP-3027

Bibliograph.
Daten
Mattes, Michael: Design and Implementation of a Framework for Online Evolution of Robotic Behaviour.
Universität Stuttgart, Fakultät Informatik, Elektrotechnik und Informationstechnik, Diplomarbeit Nr. 3027 (2010).
67 Seiten, englisch.
CR-Klassif.D.2.2 (Software Engineering Design Tools and Techniques)
Kurzfassung

The framework is written in C++ as a plug-in to the existing symbricator simula- tion environment that is an extension of delta3D simulation engine. The simulation environment provides models of existing robots complete with sensors and actua- tors, a mechanism to add controllers to the robots and arenas to deploy them to. The robot controllers use the sensor inputs to compute activation values for the actuators. The computation of the activation values for each robot is done by ar- tiï¬cial neural networks. Generating those neural networks is the main purpose of the framework. Generally they are created by an evolutionary algorithms and are evolved over a period of time to show the desired robotic behavior. The graphviz C API is used to save resulting nets to disk, to visualize them or to load existing nets from disk to use them as a starting point for evolution or to ob- serve the behavior of robots controlled by them. One of the main problems using evolutionary algorithms is evaluating the behavior shown by the robotic. This is complicated further by the principle of online evolution, which means only data available to the robot can be used for evaluation. Fitness functions assign a score to the behavior shown by the robot which is used in the process of evolution. To alleviate the problems of evaluating robotic behavior a set of ï¬tness functions suit- able for standard scenarios like wall following and food gathering will be provided by the framework. They can either be used to test and compare new evolutionary algorithms or as an example and starting point for customized ï¬tness functions. After ï¬nishing evolution within the simulation environment, the controller net- works must be transferred to real robots where they must adapt themselves to the changes in their environment. To help users with this task all necessary framework classes like classes for neural networks, evolutionary algorithms as well as robot controllers are written in such a way that permits using them within the simulation environment as well as within real hardware.

Volltext und
andere Links
PDF (4571655 Bytes)
Zugriff auf studentische Arbeiten aufgrund vorherrschender Datenschutzbestimmungen nur innerhalb der Fakultät möglich
Abteilung(en)Universität Stuttgart, Institut für Parallele und Verteilte Systeme, Bildverstehen
BetreuerSchlachter Florian
Eingabedatum7. Juli 2010
   Publ. Informatik