Automated feature selection in neuroevolution

Abstract

Feature selection is a task of great importance. Many feature selection methods have been proposed, and can be divided generally into two groups based on their dependence on the learning algorithm/classifier. Recently, a feature selection method that selects features at the same time as it evolves neural networks that use those features as inputs called Feature Selective NeuroEvolution of Augmenting Topologies (FS-NEAT) was proposed by Whiteson et al. In this paper, a novel feature selection method called Feature Deselective NeuroEvolution of Augmenting Topologies (FD-NEAT) is presented. FD-NEAT begins with fully connected inputs in its networks, and drops irrelevant or redundant inputs as evolution progresses. Herein, the performances of FD-NEAT, FS-NEAT and traditional NEAT are compared in some mathematical problems, and in a challenging race car simulator domain (RARS). On the whole, the results show that FD-NEAT significantly outperforms FS-NEAT in terms of network performance and feature selection, and evolves networks that offer the best compromise between network size and performance.

Analysis of a feature-deselective neuroevolution classifier (FD-NEAT) in a computer-aided lung nodule detection system for CT images

“…We presented a novel feature-deselective classifier, FD-NEAT in [14]. FD-NEAT is based on NEAT, which is a method that uses genetic algorithms (GAs) to evolve the topology and the weights of networks that best fit the complexity of the task at hand.…”

“…FD-NEAT's performance was previously examined on several simple feature selection experiments [14], e.g. the classical "exclusive or" classification problem and maneuvering a robotic car around a race track by selecting relevant sensors in a race car simulator environment (RARS) [28].…”

“…The networks are evolved in a population of 200 networks for 200 generations. The best network, namely the network with the highest fitness (computed using a fitness function based on the classification accuracy defined in [15] and used in [14]) on the training set at the end of the evolutionary process over ten runs is used to classify the nodules in the test set. Additional details about specific parameters that are used in the FD-NEAT algorithm are given in the Appendix and in [1].…”

“…The features that have been proposed in the literature include gradient field descriptors, invariant features, shape and regional descriptors [1][2][3][17][18]. In [14], we presented a novel feature deselection method called Feature Deselective NeuroEvolution of Augmenting Topologies (FD-NEAT). The method is a feature-deselective version of the NeuroEvolution of Augmenting Topologies (NEAT) algorithm proposed by Stanley and Miikkulainen [15].…”

Predictive feature selection for genetic policy search

“…We empirically show that PFS-NEAT is able to learn near optimal control policies in both environments even as the number of irrelevant features increases. The algorithm is also able to identify feature subsets that are comprised of higher fractions of relevant sensors as compared to FS-NEAT [55], FD-NEAT [48], and SAFS-NEAT [33], which can cause PFS-NEAT to outperform these competitors. These experiments illustrate that PFS-NEAT is an effective feature selection algorithm that is suited to scaling up genetic policy search techniques to high-dimensional problems.…”