Improving the performance and accuracy of time series modeling based on autonomic computing systems

Abstract

The challenges involved in integrating, maintaining and administrating real-world systems have motivated the proposal of the Autonomic Computing area which aims at making systems capable of self-managing their tasks, avoiding or minimizing the human interference. Autonomic Computing self-managing aspects are provided by the Autonomic Manager which relies on a structure called control loop. This loop is responsible for monitoring components, analyzing information and taking decisions to optimize, configure, heal and protect systems. Among such steps, the analysis is probably the most complex of the loop, due to it needs to model data, infer situations, estimate behavior or states, and make prediction. In order to proceed with the analysis, we need to study how systems produce information over time (here called time series) and, by modeling such information, we can understand system components behavior, transitions, relations to other systems, make estimations and predictions. In that sense, many techniques were designed to model such temporal information or time series, most of them involve statistics and artificial neural networks which require specific configurations based on series properties, what may not be bearable for online systems. All this scenario motivated this paper which proposes an approach to simplify and improve the accuracy of modeling, estimating and predicting time series. This approach considers chaos theory tools to unfold time series and organize them into multidimensional vectors which better represent interdependencies among observations. Techniques (such as artificial neural networks, e.g. RBF, RRBF, TDNN and ATNN, and statistical tools, e.g. Markov Models) can take advantage of such reorganization and, therefore, improve the accuracy of modeling, estimating and predicting. Experimental results confirm that the proposed approach provides good accuracy at very low computational costs.

Improving the accuracy of a flood forecasting model by means of machine learning and chaos theory

“…The estimated values for the maximum Lyapunov exponent are as follows: k ¼ 0:153 for March's river level time series and k ¼ 0:028 for April's river level time series. As there is k [ 0 in both cases, it can be stated that both time series have a chaotic behavior [8].…”

“…In this way, the gathered data can be interpreted as a time series, which allows us to study, model and investigate the question of behavioral prediction. A simple definition of a time series is that it is an orderly sequence of observations collected at regular intervals [7,8]. This means that the data collected by the REDE system constitutes a time series which can be studied in the light of the concepts of time series analysis.…”

“…Our paper [9] has shown that treating a time series in a varied way (i.e., using data characteristics rather than raw values) could improve the accuracy of machine learning techniques. Our reason for employing chaos theory techniques [7,8,10] is that they enable the time series to be modeled in a varied way and can thus improve the accuracy of machine learning.…”

A Stable and Online Approach to Detect Concept Drift in Data Streams

“…The parameters used to generate the data streams were: i) for the sinusoidal series, amplitude equals to 1, period equals to 40 and phase equals to 0; ii) for the Logistic map, the initial condition is 0.5 and rate is 3.8, which develops a chaotic behavior, difficulting modeling and therefore prediction using Statistical tools [19]; and iii) the sigmoid function was parameterized with smoothing factor s = 0.01 and transition point at t = 5000. After producing all data streams, every one was normalized in range [0,1].…”

Uma abordagem tolerante a falhas para a previsão de desastres naturais baseada em IoT e aprendizado de máquina

“…A área de análise de séries temporais permite estudar, modelar e prever sistemas com essas características. Séries temporais são definidas como dados organizados em termos de suas variáveis e de suas observações ao longo do tempo (ISHII; MELLO, 2012;MELLO, 2011), ou seja, uma série temporal é uma sequência de observações ordenadas no tempo e capturadas em intervalos constantes. Portanto, os dados coletados pelo sistema REDE constituem uma série temporal, o que permite analisá-los sob a luz dos conceitos dessa área.…”