Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)

First Advisor

Homaifar, Dr. Abdollah


Evolutionary algorithms work on the notion of “best fit will survive” criteria. This makes evolving a cooperative and diverse population in a competing environment via evolutionary algorithms a challenging task. Analogies to species interactions in natural ecological systems have been used to develop methods for maintaining diversity in a population. One such area that mimics species interactions in natural systems is the use of niching. Niching methods extend the application of EAs to areas that seeks to embrace multiple solutions to a given problem. The conventional fitness sharing technique has limitations when the multimodal fitness landscape has unequal peaks. Higher peaks are strong population attractors. And this technique suffers from the ‘curse of population size’ in attempting to discover all optimum points. The use of high population size makes the technique computationally complex, especially when there is a big jump in fitness values of the peaks. This work introduces a novel bio-inspired niching technique, termed Fitness Proportionate Niching (FPN), based on the analogy of finite resource model where individuals share the resource of a niche in proportion to their actual fitness. FPN makes the search algorithm unbiased to the variation in fitness values of the peaks and hence mitigates the drawbacks of conventional fitness sharing. FPN extends the global search ability of Genetic Algorithms (GAs) for evolving hierarchical cooperation in genetics-based machine learning and dynamic clustering. To this end, this work introduces FPN based resource sharing which leads to the formation of a viable default hierarchy in classifiers for the first time. It results in the co-evolution of default and exception rules, which lead to a robust and concise model description. The work also explores the feasibility and success of FPN for dynamic clustering. Unlike most other clustering techniques, FPN based clustering does not require any a priori information on the distribution of the data.