There are a lot of scenarios in which adapting or learning systems have adavantages over traditionally programmed ones: Speech recognition software uses this approach as well as stock market analysis programs or robot controls. They all have something important in common: They have to function in changing environments that can't be anticipated by the manufacturer.
During the last decades, several approaches have been developed to solve this problem. Artificial neural networks and genetic programming have adapted principles used in nature, some others like hidden Markov models or bayesian networks have their origin in statistics.
Reinforcement learning is one important method of realising machine learning without having a supervising instance. The agent tries different actions and receives a feedback that indicates if they've been good or not.
Reinforcement learning can be applied to very different scenarios. In mobile robotics, it's useful for the implementation of both low-level behaviour and high-level planning problems.
In this work, two systems are to be implemented: Obstacle avoidance as a quite simple topic in mobile robotics, and the decision when to stop exploration to return to a charging station as a problem on a higher level.
Since most robotic systems need a map to localize themselves and plan their actions, a map building algorithm will be implemented as an additional task.
The chapters 2 and 3 give a short overview over the theoretical background and explain the reinforcement learning problem, main solution methods and cartography methods based on a laser scanner. Afterwards, the simulation environment used for the realisation (chapter 4) and the implementation of the algorithms itself are described (chapter 5). In the last section, the results are presented and evaluated.