Divergence-based Bit Allocation for Indoor Environment Classification

Machine learning is rapidly gaining attention for enabling cognitive functionalities in future networks. Machine learning units (MLU) extract the hidden knowledge in data to perform tasks, and their input attributes have various levels of relevancy for MLU outcome. Recently, a novel quantization bit allocation was introduced in which Kullback-Leibler divergence (KLD) is used to consider this relevancy. It has been shown that KLD bit allocation achieves high performance gains when applied to the cart inverted pendulum problem. In this paper, we further evaluate and provide insights into advantages of employing this approach instead of its conventional alternatives, a method using sum of squared errors (SSE) and equal sharing, for indoor environment classification with real measurements. This study covers both scalar and vector quantization, and different hypotheses: neural network, decision tree, 1-nearest neighbor and support vector machine. Furthermore, MLU performance and its robustness to packet drop with selected bit allocations of KLD and SSE is evaluated. Simulations show that classification accuracy achieved by the KLD approach is significantly higher or in a few particular cases similar to the conventional methods, and it does not affect robustness of the MLU regarding packet loss.