Section: New Results

Learning for an efficient and dynamic management of network resources and services

Learning in networks

Participants : Dana Marinca, Nesrine Ben Hassine, Pascale Minet, Selma Boumerdassi.

This work is a joint work with Dominique Barth (University of Versailles-Saint-Quentin). To guarantee an efficient and dynamic management of network resources and services we intend to use a powerful mathematical tool: prediction and learning from prediction. Prediction will be concerned with guessing the evolution of network or network components state, based on knowledge about the past elements and/or other available information. Basically, the prediction problem could be formulated as follows: a forecaster observes the values of one or several metrics giving indications about the network state (generally speaking the network represents the environment). At each time $t$, before the environment reveals the new metric values, the forecaster predicts the new values based on previous observations. Contrary to classical methods where the environment evolution is characterized by stochastic process, we suppose that the environment evolution follows an unspecified mechanism, which could be deterministic, stochastic, or even adaptive to a given behavior. The prediction process should adapt to unpredictable network state changes due to its non-stationary nature. To properly address the adaptivity challenge, a special type of forecasters is used: the experts. These experts analyse the previous environment values, apply their own computation and make their own prediction. The experts predictions are given to the forecaster before the next environment values are revealed. The forecaster can then make its own prediction depending on the experts' "advice". The risk of a prediction may be defined as the value of a loss function measuring the discrepancy between the predicted value and the real environment value. The principal notion to optimize the behavior of the forecasters is the regret, seen as a difference between the forecaster's accumulated loss and that of each expert. To optimize the prediction process means to construct a forecasting strategy that guarantees a small loss with respect to defined experts. Adaptability of the forecaster is reflected in the manner in which it is able to follow the better expert according to the context.

Our purpose is to apply on-line learning strategies to:

• Wireless Sensor Networks (WSNs) to predict the quality of a wireless link in a WSN, based on the LQI metric for instance and take advantage of wireless links with the best possible quality to improve the packet delivery rate. We model this problem as a forecaster prediction game based on the advice of several experts. The forecaster learns on-line how to adjust its prediction to better fit the environment metric values. A forecaster estimates the LQI value using the advice of experts.

• Content Delivery Networks (CDNs) to predict the number of solicitations of video contents to cache the contents with the highest popularity.

• Data centers require a huge amount of energy. As an example, in 2014, the electric consumption of all date centers will be larger than 42 TWh, and after 2020 the CO2 production will be larger then 1.27 GTons, ie. more than the aeronautic industry (GeSI SMARTer 2020 report). These "frightening" figures led the research community to work on the management of energy consumption. Several tracks have been explored, among which the optimization of computation and load balancing of servers. At present, we work on tools dedicated to traffic prediction, thus allowing a better management of servers. Our work consists in modeling the traffic specific to data centers and apply different statistical prediction methods.

Tools for learning and prediction

Participants : Dana Marinca, Nesrine Ben Hassine, Pascale Minet.

In 2015, Nesrine Ben Hassine developped an extraction tool to provide real traces from YouTube. these real traces are used as a learning sample by the different prediction algorithms used.

Nesrine Ben Hassine and Dana Marinca extended their simulation tool developed in Python to integrate:

• various prediction strategies SES (Single Explonential Smoothing), DES (Double Exponential Smoothing), Basic and enhanced basic, strategies based on averages (e.g. Average on a Moving Window), regressions (e.g. polynomial or Savitzky Golay), as well as prediction strategies adapting dynamically their parameters according to the loss obtained.

• various loss functions (e.g. absolute value, square). The prediction accuracy is evaluated by a loss function as the discrepancy between the prediction value and the real number obtained.

• different forecaster strategies: Best expert, Exponential Weighted Average, K Best-Experts, etc.

With these tools, we can now tune parameters of prediction strategies and evaluate them.

Popularity prediction in CDNs

Participants : Dana Marinca, Nesrine Ben Hassine, Pascale Minet.

To predict the popularity of video contents, expressed as the number of solicitations, we compared three prediction strategies: Single Exponential Smooting (SES), Double Exponential Smoothing (DES) and Basic. The best tuning of each strategy is determined, depending on the considered phase of the solicitation curve. For DES, values of the smoothing factor close to 1 probide the best results. We study the behavior of each strategy within a phase and around a phase change, where a phase is defined as an interval of time during which a measured metric remains relatively stable.

Basic expert makes large errors at the phase change, but it quickly corrects its prediction and it is the expert having the closest prediction to the real value within a phase. DES expert provides also good quality predictions within a phase. Since DES and Basic experts outperform the SES expert, we recommend the use of on the one hand, the best DES expert per phase within a phase and on the other hand, the Basic expert to automatically detect phase changes, because of its better reactivity. This self-learning and prediction method can be applied to optimize resources allocation in service oriented architectures and self-adaptive networks, more precisely for cache management in CDNs.

Automatic phase detection in popularity evolution of video contents

Participants : Dana Marinca, Nesrine Ben Hassine, Pascale Minet.

In Content Delivery Networks (CDNs) where experts predict the number of solicitations of video contents, simulations based on real YouTube traces show that the accuracy of prediction is improved by splitting the video content profile in contiguous phases. A phase is an interval of time during which a measured metric remains relatively stable. The best expert per phase outperforms the best expert on the whole video content profile. Different prediction methods are compared and also different phase change-points detection methods are evaluated:

• the R tool using Bayesian inference,

• the Basic expert (an important loss may indicate a phase change),

• a fixed time interval (e.g. each week).

The goal is to identify the method (or method parameters) minimizing the cumulated discrepancy compared to real solicitations of video contents. The use of this machine learning method allows the Content Delivery Network to self-adapt to users solicitations by caching the most popular contents near the end users. More generally, such method can be applied to decide which contents should be replicated to improve the performance of audio and video applications and maximize the satisfaction degree of users.