Vorgelegte Bachelor-Thesis von Felix End aus Bad Homburg

Hiermit versichere ich, die vorliegende Bachelor-Thesis ohne Hilfe Dritter nur mit den angegebenen Quellen und Hilfsmitteln angefertigt zu haben. Alle Stellen, die aus Quellen entnommen wurden, sind als solche kenntlich gemacht. Diese Arbeit hat in gleicher oder ähnlicher Form noch keiner Prüfungs-behörde vorgelegen. Darmstadt, den 30. September 2015 (Felix End) Damage or wear of robot parts or changes in the environment can cause the previously optimal behavior of a robot to become inadequate. Having backup solutions in such a situation can allow the robot to continue doing its task. The majority of Reinforcement Learning (RL) algorithms is not designed to find backup solutions. This thesis presents a new hierarchical RL algorithm that builds on the ideas of the Hierarchical Relative Entropy Policy Search (HiREPS) algorithm to learn multiple solutions. The policy that is learned is a ’mixed options ’ policy, which consists of a gating policy and option policies. The gating policy selects the option policy, which then chooses an action. The different option policies represent different solutions. The presented algorithm mimics the two level structure of the ’mixed options ’ policy. It has an upper level which optimizes the gating and a lower level which optimizes the option policies. Both the upper and the lower level optimization are extensions of Relative Entropy Policy Search (REPS). Subsequently, the whole algorithm belongs to the group of policy search algorithms. The experiments show that the new algorithm finds more versatile solutions than HiREPS and its best solution is in quality similar to that of HiREPS. Zusammenfassun