Zusammenfassung:
Artifi cial neural networks are an approach towards creating arti cial intelligence, loosely modelled after the human brain and rst introduced by Frank Rosenblatt in 1962. Four years earlier Friedberg introduced genetic algorithms, which mimic biological evolution to optimize computer programs. Arti cial neural networks were rediscovered in 1986 and later became widely accepted. Nowadays they are successfully applied in a wide variety of commercial and scienti c elds. But even after many decades of research, applying neural networks to real world problems remains a challenge. This thesis is about the combination of arti cial neural networks and genetic algorithms, which is called neuroevolution. This concept has been applied to many real world problems, but this thesis focuses on the control of robots. Whereas most research papers only present the algorithms producing the desired results, this thesis systematically compares di erent approaches for the control of autonomous robots. To accomplish this, a foraging task along with an appropriate environment is designed. For each experiment, a population of robots is placed within the environment. Each robot has the goal to survive as long as possible and to collect as many resources as possible. The main algorithm, which is evaluated in this thesis is based on neuroevolution. To assess the performance, two additional robot controllers were designed. The rst controller steers the robots randomly through the environment without any goal-oriented behavior and is supposed to delimit the lower bound for the achievable performance. The second algorithm is hand crafted and especially tuned for the de ned task. It is apparent that this algorithm performs close to the theoretical optimum and therefore can be used to estimate the upper bound for the achievable performance. In order to conduct these experiments in an ecient manner, the BRAIn framework was designed and implemented. The algorithms were tested on the marXbot robotic platform. The environment, as well as the robots were simulated using the ARGoS simulator. The results show that it is still dicult and tedious to apply arti cial neural networks to robotics. However, the thesis also demonstrates some key benefi ts of this concept. The thesis is concluded with an overview of the advantages and disadvantages of neural networks and a comparison with other approaches.
Download:
thesis-wuertinger_2011-12-19.pdf