A Reward and Penalty Based Approach for Online Feature Selection

Abstract

For realization of self-adaptive systems, the underlying model needs to update its structure and parameters autonomously with the change in the environment. For such systems the model structure and parameters are learnt using data, and selection of features plays an important role in the learning process and in turn in the performance of the system. In this paper, we focus on the problem of autonomous selection of features that are suitable for a given environment. A simple but effective approach is presented in this paper that is based on the concept of reward and penalty. To test the efficacy of the approach, it is implemented in a mobile robot. The robot needs to classify human from other objects in an environment using a set of sensors and support vector machine-based classifier. As the robot moves from one environment to the other, the set of features that were used initially might not be suitable for the new environment. With the proposed approach, the set of suitable features are identified on the fly for the new environment and a low false positive rate is achieved. Further, the approach will also be suitable for other applications based on Internet of Things that attempt to realize smart environments by embedding numerous sensors in the environment. The aim of such applications is to perform desired task with reduced human intervention. These applications, typically running in resource constrained devices, demand for processing of streaming data with constrained memory space and time. Thereby, any redundancy in features can increase the computational time as well as consume memory space.