Assessment of Predictive Ability of Artificial Neural Networks Using Holographic Mapping

In this study, artificial neural networks (ANNs) were used to reveal a quantitative relationship between catalytic composition and catalytic activity. This relationship was predefined using a hypothetical experimental space described by a multidimensional polynomial. The predictive ability of ANNs was investigated, i.e. an attempt was done to evaluate how ANNs can envisage a given hypothetical experimental space. Data sets for training, validation and testing of ANNs were obtained from the hypothetical experimental space using two different ways of sampling. Data were selected, (i) by means of our optimization algorithm called Holographic Research Strategy (HRS); and (ii) randomly. In order to model real experimentation, data were also generated with error. The relationship between the complexity of different network topologies and their predictive ability was investigated. It was shown that when data used for training have been perturbed with a given level of noise, less complex network architectures give acceptable accuracy. Additionally, estimated experimental spaces were visualized in a 2D layout by means of Holographic Mappings (HMs). Analysis of HMs revealed that ANNs trained by data sets obtained upon an optimization procedure provides better description of the experimental space in the vicinity of the optimum than ANNs trained by randomly selected data sets. This fact indicates again the importance of the optimization in combinatorial catalyst library design.