Multilayer Perceptron: NSGA II for a New Multi-objective Learning Method for Training and Model Complexity

Senhaji, Kaoutar; Ramchoun, Hassan; Ettaouil, Mohamed

doi:10.1007/978-3-319-91337-7_15

Cited by 5 publications

(1 citation statement)

References 17 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…During selection stage of evolution, population of generated networks will now be evaluated not only for accuracy but for multiple objectives which are none other than the performance metrics. Recently, NSGAII (Non-dominated Sorting Genetic Algorithm), a popular evolutionary algorithm (EA), has been used to perform multi-objective optimization of neural networks minimizing the perceptron error and the network complexity [45]. However, applications of EAs to optimize neural networks have been limited to the variation in network hyper-parameters such as network architecture, number of neurons and parameters namely; connection weights [45]- [50].…”

Section: Neuroevolutionmentioning

confidence: 99%

Analytics of Heterogeneous Breast Cancer Data Using Neuroevolution

et al. 2019

View full text Add to dashboard Cite

Breast cancer prognostic modeling is difficult since it is governed by many diverse factors. Given the low median survival and large scale breast cancer data, which comes from high throughput technology, the accurate and reliable prognosis of breast cancer is becoming increasingly difficult. While accurate and timely prognosis may save many patients from going through painful and expensive treatments, it may also help oncologists in managing the disease more efficiently and effectively. Data analytics augmented by machine-learning algorithms have been proposed in past for breast cancer prognosis; and however, most of these could not perform well owing to the heterogeneous nature of available data and model interpretability related issues. A robust prognostic modeling approach is proposed here whereby a Pareto optimal set of deep neural networks (DNNs) exhibiting equally good performance metrics is obtained. The set of DNNs is initialized and their hyperparameters are optimized using the evolutionary algorithm, NSGAIII. The final DNN model is selected from the Pareto optimal set of many DNNs using a fuzzy inferencing approach. Contrary to using DNNs as the black box, the proposed scheme allows understanding how various performance metrics (such as accuracy, sensitivity, F1, and so on) change with changes in hyperparameters. This enhanced interpretability can be further used to improve or modify the behavior of DNNs. The heterogeneous breast cancer database requires preprocessing for better interpretation of categorical variables in order to improve prognosis from classifiers. Furthermore, we propose to use a neural networkbased entity-embedding method for categorical features with high cardinality. This approach can provide a vector representation of categorical features in multidimensional space with enhanced interpretability. It is shown with evidence that DNNs optimized using evolutionary algorithms exhibit improved performance over other classifiers mentioned in this paper. INDEX TERMS Breast cancer prognostic modelling, entity embedding, deep learning networks, evolutionary algorithms, fuzzy inferencing.

show abstract

Section: Neuroevolutionmentioning

confidence: 99%