Artificial intelligence in mobile autonomous sensor networks

Abstract

This work is concerned with a common and frequent engineering task of constructing a least-squares model based on measured data. What is provided are data from one or more sensors. Also is known that an unknown model is generated by set of functions that come from a known family of function. However it is unknown ahead of time what subset of functions and what concrete parameters generated the data. The task is to propose a method of obtaining this model using minimal number of measurements. This method also provides rules for determination of numerical values of variance and error that let us make a decision whether the number of measurements is sufficient, or additional measurements need to be done, thus the variance of the model is used as the stopping criterion in this case. The proposed analysis is relevant in practical situations where consistency theorems do not give necessary guidance for small samples of data.

The method of least-squares provides a model that can be accurate (small total error). This model is optimal in the least-squares sense, but the variance of the model is not guaranteed and can be large in certain sub-domain. In case of large data samples consistency theorems guarantee convergence to the true model. However, small data samples can result in a catastrophic model. To avoid this kind of models additional measurements are required. The purpose of these measurements is to decrease the model variance that is indicative of a misleading model. Use of a minimization procedure that tests potential future measurement points and selects an optimal or sub-optimal data point with a good impact on the model variance is proposed. The non-linear parameters of the model are calculated by genetic algorithm.

Part of this work is also an algorithm for searching the maximum admissible variance for a group of models and implementation of algorithms in MATLAB and C#. For practical verification wireless control of an autonomous robot based on image recognition was elaborated.