Optimization of Multi-Layer Perceptron Deep Neural Networks using Genetic Algorithms for Hand Gesture Recognition

Applications of wearable sensors for Hand Gesture Recognition (HGR) have been gaining popularity in recent years. Among the proposed methods, deep neural networks with many hidden layers are promising to address the requirements of this wearable activity recognition. They can directly uncover features tied to the dynamics of HGR, from simple motion encoding in lower layers to more complex motion dynamics in upper layers. However, these methods require many efforts of researches to build an efficient neural network architecture. This study proposes an integrated method that allows finding the best neural networks for HGR using wearable sensors. The proposed method consists of two parts: (i) A generic Multi-Layer Perceptron (MLP) deep neural network and (ii) A genetic algorithm. We applied the genetic algorithm to find the best network architecture in terms of accuracy. At each generation of the algorithm, a new set of architecture was created with different Hyper parameters (the activation, optimizer, the number of layers, neurons and epochs). Extensive experiments were conducted on a dataset containing 18.000 gesture samples from 20 subjects. Experimental results demonstrated the performance and efficiency of the proposed methods in finding deep neural network architectures for HGR. The obtained neural network achieves 89.21% of accuracy and outperforms the previous study on the same dataset.