Quality assured efficient engineering of feedforward neural net-works with supervised learning (QUEEN) evaluated with the 'pima Indians diabetes database'

摘要

The QUEEN method is based on four main concepts: 1. The QUEEN phase model is derived from the spiral model of Bohm and integrates the development of a neural network. 2. An overall strategy for the development process enables a continuous supervision, assessment and quality assurance of each step, from the collection of the examples to the evaluation of the constructed neural network. For the assessment of the quality achieved in the development process, a novel quality indicator is introduced. This indicator gives a measure of the complexity of a task in a given representation. This strategy of QUEEN involves the stepwise simplification of the task. 3. The development of the neural networks is structured by the definition of an order over neural networks. The order takes into account the complexity of the interpretation of the neural network by an expert of the application domain. To yield easily interpretable neural networks, and also to get simple models that enable the detection of data artefacts, the development is started with the simplest adequate neural network. 4. The developer is provided with a set of diagnostic methods and tools that will identify and eliminate reasons for difficulties. The novel quality indicator e.g. provides the developer with a diagnostic tool that will identify situations where a representation or a network is unnecessarily complex. QUEEN was developed and successfully evaluated in more than 20 projects mostly in the medical application domain. This paper presents the concepts of QUEEN and describes how QUEEN was applied to set-up a feedforward neural network on the pima indians diabetes database, a database that has been used as a benchmark in several studies. QUEEN highlighted several severe data artefacts in this database.