Realtime mobile bandwidth prediction using LSTM neural network and Bayesian fusion

Abstract

With the increasing popularity of mobile Internet and the higher bandwidth requirement of mobile applications, user Quality of Experience (QoE) is particularly important. For applications requiring high bandwidth and low delay, such as video streaming, video conferencing, and online gaming, etc., if the future bandwidth can be estimated in advance, applications can leverage the estimation to adjust their data transmission strategies and significantly improve the user QoE. In this paper, we focus on accurate bandwidth prediction to improve user QoE. Specifically, We study realtime mobile bandwidth prediction in various mobile networking scenarios, such as subway and bus rides along different routes. The primary method used is Long Short Term Memory (LSTM) recurrent neural network. In individual scenarios, LSTM significantly improves the prediction accuracy of state-of-the-art prediction algorithms, such as Recursive Least Squares (RLS) by 12% in Root Mean Square Error (RMSE) and by 17% in Mean Average Error (MAE). We further developed Multi-Scale Entropy (MSE) to analyze the bandwidth patterns in different mobility scenarios and discuss its connection to the achieved accuracy. For practical applications, we developed Model Switching and Bayes Model Fusion to use pre-trained LSTM models for online realtime bandwidth prediction.

Introduction

We have witnessed the tremendous growth of mobile Internet traffic in the recent years. Users are increasingly spending more time on mobile apps and consuming more content on their mobile devices. The mobile traffic growth is expected to accelerate in the foreseeable future with the introduction of 5G wireless access and new media-rich applications, such as Virtual Reality and Augmented Reality. Providing better Quality of Experience (QoE) to mobile app users is of great significance. However, one main challenge for mobile app developers and content providers is the high volatility of mobile wireless connections. The physical channel quality of a mobile user is constantly affected by interference generated by other users, his/her own mobility, and signal blockages from static and dynamic blockers [1], [2]. The bandwidth available for a mobile session is ultimately determined by the adaptations cross the protocol stack, ranging from adaptive coding and modulation at PHY layer, cellular scheduling at data link layer, hand-overs between base stations, to TCP congestion control, etc. For many mobile apps involving user interactivity and/or multimedia content, e.g., gaming, conferencing, and video streaming, it is critical to accurately estimate the available bandwidth in realtime to deliver a high quality of user Quality-of-Experience (QoE). In the example of video streaming, many recent algorithms on Dynamic Adaptive Streaming over HTTP (DASH) optimize the video rate selection for upcoming video chunks based on the predicted TCP throughput in a future time window of several seconds [3], [4], [5]. If TCP throughput in the next few seconds can be accurately predicted, the selected video rate can maximize the delivery video quality and avoid video freeze. Interactive video conferencing has much tighter delay constraint than streaming. To avoid self-congestion, the available bandwidth on cellular links has to be accurately estimated in realtime, which is used to guide the realtime video coding and transmission strategies [6], [7]. Bandwidth overestimate will lead to long end-to-end video delay or freezing, and bandwidth underestimate will lead to unnecessarily poor perceptual video quality. Again, accurate realtime bandwidth prediction is crucial for delivering a good conferencing experience, especially in mobile networking scenarios.

In this paper, we study realtime mobile bandwidth prediction using Long Short Term Memory (LSTM) [8] recurrent neural network and Bayes fusion. Recent advances in Deep Learning have demonstrated that Recurrent Neural Networks (RNN) are powerful tools for sequence modeling and can learn temporal patterns in sequential data. RNNs have been widely used in Natural Language Processing (NLP), speech recognition, and time series processing [9], [10]. There are rich structures in realtime mobile network bandwidth evolution, due to user mobility patterns, wireless signal propagation laws, physical blockage models, and the well-defined behaviors of network protocols. This presents abundant opportunities for developing LSTM-based realtime mobile bandwidth estimation. The main idea is to offline train LSTM RNN models that capture the temporal patterns in various mobile networking scenarios. The trained LSTM RNN models will be used online to predict in realtime the network bandwidth within a short future time window. Specifically, we investigate the following research questions:

• How much prediction accuracy improvement can LSTM Deep Learning models bring over the conventional statistical prediction models?

• How predictable is realtime bandwidth at different prediction intervals under different mobility scenarios? Is the LSTM prediction accuracy dependent on specific mobility scenarios?

• Should one train a separate LSTM model for each mobility scenario, or train a universal LSTM model that can be used in different scenarios?

• How should one switch between different pre-trained models when user mobility pattern changes? Is it possible to fuse the predictions from multiple models to adapt to mobility pattern changes?

Towards answering these questions, we made the following contributions:

• We conducted a mobile bandwidth measurement campaign to collect consecutive bandwidth traces in New York City. Our traces cover different transportation methods along different routes at different times of the day.¹ The traces are described in detail in Section 4.

• We developed LSTM models for realtime one-second ahead and multi-second ahead bandwidth predictions. Through extensive experiments on our own dataset and the HSDPA dataset [11], we demonstrated that LSTM significantly outperforms the existing realtime bandwidth prediction algorithms. Our LSTM models and their performance are presented in Sections 3 LSTM based realtime bandwidth prediction, 4 Data collection and performance evaluation.

• We systematically evaluated the sensitivity of LSTM models to different mobility scenarios by comparing the accuracy of per-scenario, cross-scenario and universal predictions. Using Multi-Scale Entropy (MSE) analysis, we studied the connection between prediction accuracy and bandwidth regularity at different time scales. MSE also provides us with guidelines to explore cross-scenario bandwidth prediction. The analysis is presented in Section 5.

• We designed model switching, which selects the best-performing LSTM model in recent history, and can converge to the model that fits the most to the current mobility scenario. We also designed model fusion, which generates predictions through Bayesian fusion of outputs of multiple models, and can make smooth transition from one mobility scenario to another. Model switching and fusion are presented in Section 6.

The rest of the paper is organized as the following. The related work on realtime bandwidth prediction is reviewed in Section 2. We formally define the realtime bandwidth prediction problem and introduce our LSTM based prediction models in Section 3. The performance of LSTM models is evaluated by a public dataset and our own dataset in Section 4. We conduct Multi-Scale Entropy analysis on our collected bandwidth traces and analyze the prediction accuracy in Section 5. In Section 6, we present Model Switching and Model Fusion to address mobility pattern changes. The paper is concluded with future work in Section 7.