In all fields of pattern recognition, various techniques are available to classify instances of patterns, each approach being characterized by its own virtues and shortcomings. The idea of combining the output of multiple classifiers has been studied for several years with the expectation that when compared to the single best classifier available, ensembles of classifiers may yield superior recognition capabilities. Multiple classifier combination is an example of the more general and fundamental problem of integration of information from multiple sources. How to choose a suitable combination algorithm is still a problem and this is a topic of research in this project.
Given a set of ranked classifier hypothesis, voting techniques which are well-known from, e.g., economics or politics can be applied to yield a combined outcome. The first step in this project will be to investigate what performance gains can be achieved with these relatively simple algorithms.
A more ambitious system for classifier combination is the multi-agent paradigm. The multi-agent paradigm uses agents, autonomous software components that can be designed, implemented, trained and operated independently, but that work together using a mixed quantitative and symbolic negotiation protocol. A system of such agents is dynamic and flexible in varying conditions and furthermore provides a powerful method to represent highly non-linear metaclass boundaries.
The main goal of this PhD project is to research the multi-agent alternative for combining classification methods. This work should result in a PhD-thesis, an agent negotiation protocol specialized for pattern recognition and at least one such multi-agent system in a pattern recognition application: handwriting recognition.