Dissertations, Theses, and Capstone Projects

Date of Degree


Document Type


Degree Name



Computer Science


Elizabeth Sklar

Committee Members

Katherine St. John

Martijn Schut

Ioannis Stamos

Subject Categories

Artificial Intelligence and Robotics | Computer Sciences


machine learning, genetic programming, artificial life, multiagent systems, robotics, evolutionary systems


In the Evolutionary Computation (EC) research community, a major concern is maintaining optimal levels of population diversity. In the Multiagent Systems (MAS) research community, a major concern is implementing effective agent coordination through various interaction mechanisms. These two concerns coincide when one is faced with Evolutionary Multiagent Systems (EMAS).

This thesis demonstrates a methodology to study the relationship between interaction mechanisms, population diversity, and performance of an evolving multiagent system in a dynamic, real-time, and asynchronous environment. An open sourced extensible experimentation platform is developed that allows plug-ins for evolutionary models, interaction mechanisms, and genotypical encoding schemes beyond the one used to run experiments. Moreover, the platform is designed to scale arbitrarily large number of parallel experiments in multi-core clustered environments.

The main contribution of this thesis is better understanding of the role played by population diversity and interaction mechanisms in the evolution of multiagent systems. First, it is shown, through carefully planned experiments in three different evolutionary models, that both interaction mechanisms and population diversity have a statistically significant impact on performance in a system of evolutionary agents coordinating to achieve a shared goal of completing problems in sequential task domains. Second, it is experimentally verified that, in the sequential task domain, a larger heterogeneous population of limited-capability agents will evolve to perform better than a smaller homogeneous population of full-capability agents, and performance is influenced by the ways in which the agents interact. Finally, two novel trait-based population diversity levels are described and are shown to be effective in their applicability.