Zusammenfassung:
In the research eld of Swarm Robotics, engineers have to decide how to organize their robotic swarm. In this context, one of the most important decisions is which robot should execute which task in order to achieve a given global mission. The corresponding optimization problem is called Task Allocation and can be tackled by various mechanisms, e.g. auctions, which are based on research in game theory. This thesis gives an overview of mechanisms for Task Allocation in Swarm Robotics. For this purpose, a new taxonomy is proposed that divides solution strategies into Heteronomous Task Allocation, Autonomous Task Allocation and Hybrid Task Allocation. The experimental part of this thesis uses this system to develop solution methods for a concrete mission that is inspired by nature: similar to a swarm of bees, multiple robots are attached to a nest that is used to deposit food. Besides foraging, the articial swarm needs to keep the nest’s temperature close to an optimal value. Otherwise gathered food cannot be processed to energy. This thesis focuses on one probabilistic, motivation-based approach and variants of centralized and decentralized reinforcement learning, which is a kind of machine learning that emphasizes the selection of actions under the observation of rewards. For comparison of the approaches, corresponding swarms are simulated in static and dynamic environments. In this context, static means that the aerial temperature and the food density are fixed. For the execution of these experiments, the Swarmulator was developed, a simulation platform that is well suited for testing mechanisms for Task Allocation in Swarm Robotics. The Swarmulator features a modular design and batch processing. Statistical data is saved in tabular form and serves as a basis for the concluding analysis of the proposed solution methods.
Download:
burger_2012_diploma-thesis.pdf