Nowadays Multi Agent Systems are applied for many problems as decentrally organized systems with agents acting autonomously and in parallel. In many cases it is desired to find capabilities like adaptivity, self-healing, controllability and reliability in such systems.

The words swarm intelligence, emergence and synergy are often used to describe the phenomenon of the development of a complex global system behavior with only locally acting and possibly simple subsystems. A systematic approach of designing local agents in a way such that a global goal is reached by guaranteeing reliability, controllability etc. at the same time, is not a trivial task. But the implementation of a Multi Agent System can still be an improvement compared to a single complex processing unit, if, due to the exploitation of parallelism and the simplicity of the agents, the performance can be increased and/or the need of resources can be reduced. E.\,g., Multi Agent Systems can be implemented for image processing in camera sensor chips or for adaptive routing algorithms in networks on a chip. Thus, in spite of the difficulties of the designing process of an emergent behavior, it might be worth doing it.

The general goal of this work is on one hand the development of a Multi Agent System model, which is suitable for an arbitrary agent control structure (agent behavior control) and on the other hand the design of techniques for the optimization of the control structure.

Cellular Automata are generally well applicable as a model for Multi Agent Systems with a spatially discrete structure and which are supposed to be implemented in hardware chips. In this thesis, a general model for Multi Agent Systems in Cellular Automata is developed, which can incorporate an arbitrary structure for the programming of the agent behavior.

Three example applications of Multi Agent Systems are defined in this thesis: the complete exploration of an unknown area with obstacles, that is divided in small sections, the complete distribution of initially mutually exclusively distributed information and the transport (routing) of messages in a two-dimensional mesh network. Agents that are controlled by a finite state machine are used in these examples. The behavior of the agents (i.\,e., the content of the finite state machines) is evolved with Genetic Algorithms, in order to reach the previously defined goals. Additionally, optimization techniques for the structural assembly of the agents are designed. Furthermore, methods for the adjustment of the configuration of the Cellular Automaton are presented. All of the methods and techniques are suggested with the intention to increase the effectivity and the efficiency of the Multi Agent Systems as a whole.

Eventually, experiments and results are presented, which evaluate the effectivity and efficiency of the Genetic Algorithms, the optimization techniques for the agents' structure as well as the methods for the adjustment of the Cellular Automaton. The Genetic Algorithms perform well in any of the example systems and find feasible solutions (agent behaviors). Not all of the adjustments and optimization techniques are effective and efficient. Particularly those methods that induce a spatial or temporal heterogeneity or that use a certain amount of randomness in the agents' behavior, are effective.