Activity recognition using back-propagation algorithm and minimum redundancy feature selection method Nadia OUKRICH* 1 , Abdelilah MAACH 1 , ElMehdi SABRI 1 . ElMahdi Mabrouk 1 1 Computer Laboratory Research and Development 1 Mohammedia Engineering School *nadiaoukrich@gmail.com Kevin BOUCHARD Center for SMART Health University of California Los Angeles United States Abstract— In this paper, we use multilayer Perceptron model and a supervised learning technique called backpropagation to train a neural network in order to recognize human activity inside smart home, and select useful features according to minimum redundancy maximum relevance. The results show that different feature datasets and different number of neurons of hidden layer of neural network yield different activity recognition accuracy. The selection of suitable feature datasets increases the activity recognition accuracy and reduces the time of execution. Furthermore, neural network using back-propagation algorithm and multilayer Perceptron model has relatively better human activity recognition performances. Keywords— activity recognition; smart home; Multilayer perceptron; back-propagation; feature selection; mutual information. I. INTRODUCTION With the development of various technologies, people want to make their lives easier and enjoyable. Smart home is a house equipped with automated systems such as lightning and door operation or equipped with any technological appliance in a house that could automate simple tasks. However, nowadays almost any electrical house components can be included in the system [1], Smart home is used for several purposes. It can improve comfort at home, enable automation of household chores and reduce energy consumption. It can provide a better- quality of life inside home by adapting behaviors of inhabitants to their preferences [2]. In fact, this last application is based on recognition of human activity inside home. Human activity recognition (HAR) intends to observe human-related actions in order to obtain an understanding of what type of activities/routines individuals perform within a time interval in order to provide a useful feedback by the system [3]. Through this paper, we referred to the concept of Activity of Daily Living (ADL), ADL has been first described by the Dr. Katz [4] as the set of activities that an individual performs in his routine to take care of himself. That includes activities such as preparing meals, getting dressed, toileting himself, etc. In general, ADL is any activity that a normal person is supposed to be able to realize in his home. Researchers distinguishes two different types of ADLs: Basic ADLs that response to primary needs of a person, these activities are composed of only a few steps and do not require real planning. Instrumental ADLs this kind of activity needs basic planning to be performed and implies objects manipulations. In this paper, we use multilayer Perceptron model, a type of Artificial Neural Network, the computation is performed using a set of simple units with weighted connections between them. Furthermore, Back Propagation (BP) algorithm [5,6] is used to set the values of the weights. Otherwise BP is a common method of training artificial neural networks in supervised learning method which calculates the gradient of a loss function with respect to all the weights in the network. The gradient is fed to the optimization method which in turn uses it to update the weights, in an attempt to minimize the loss function. We use also, a feature selection algorithm called, minimal-redundancy- maximal-relevance criterion (mRMR) [7,8] to obtain a set of subsets of features with high class-feature mutual information classification and the less feature-feature mutual information. The rest of the paper is organized as follows. Section 2 reviews related work. Section 3 describes the designing of multilayer perceptron network using BP algorithm applied to recognize human activities, the smart apartment testbed and data collection. Section 4 introduce and applicate the minimum redundancy — maximum relevance (MRMR) optimization approach method for features selection. Section 5 resume the test and results. Finally, Section 6 concludes. II. RELATED WORKS Researchers classify activity recognition approaches into two categories. The first is based on the use of visual sensing facilities, example: camera, and exploit computer vision techniques to analyze visual observations for pattern recognition [9,10]. The second category is based on the use of emerging sensor network technologies and using data mining and machine learning techniques to analyze sensors data and determine user’s behavior [11-18]. Sensors can be wearable [11] or fixed in doors or spatial place at home. Due to low cost and low power consumption, sensors based approach became a center of interest at the last decade, researchers have commonly tested the machine learning algorithms such as knowledge-driven approach (KDA) [12], evolutionary ensembles model (EEM) [13], support vector machine (SVM) [14], Naïve Bayes (NB) classifier [15], hidden Markov model (HMM) [16], and conditional random fields (CRF) [17]. While Neural networks algorithm, used in this paper, were first published in 1960. In the years following, many new techniques have been developed in the field of neural networks, and the discipline is growing rapidly [18]. Neural network has proven successful in different 978-1-5090-0751-6/16/$31.00 ©2016 IEEE 818