SMiLer: A Semi-Lazy Time Series Prediction System for Sensors

1. Introduction

SMiLer is a SeMiLazy time series prediction system for sensors. The overall framework of SMiLer comprising of two main steps: search step and prediction step. More details about it can be found in the paper:

****Jingbo Zhou, Anthony K. H. Tung; "SMiLer: A Semi-Lazy Time Series Prediction System for Sensors"; Proc. of 2015 ACM Int. Conf. on Management of Data (SIGMOD 2015)

2. Quick Run

You can see a running example in the folder "src/demo". The demo also gives a brief introudction about the API and parameters.

3. Usage

This section shows important functional APIs of SMiLer.

3.1 Data loading function

	void conf_DataEngine(vector<vector<float> >& in_bladeData_vec, 
				int sc_band, int winDim,int maxOffset)

This function is in src/smilerManager/TSManagerOnGpuIdx.h. It is the main function to load the sensor time series data into the GPU. The parameters of conf_DataEngine() are as follows:

3.1.1 Parameters of conf_DataEngine():

 'in_bladeData_vec' -- the time series of sensors, one blade is for one sensor
 'sc_band' -- the warping width for LB_keogh of DTW
 'winDim' -- the window length
 'maxOffset' -- do not load maxOffset data into (or say left them out of) GPU to avoid exceeding the bounding when query the reference time series

3.2 Continuous query function

	void TSPred_continuous_onGPUIdx(
			vector<vector<float> >& groupQuery_vec, 
			int contPrdStep, 
			vector<int>& groupQuery_blade_map,
			int y_offset, int pred_selector, 
			bool weightedTrn, bool selfCorr)

This function is in src/smilerManager/TSManagerOnGpuIdx.h. It is the main function to start the continuous query on SMiLer. By this function, SMiLer makes continuous prediction for each vector in groupQuery_vec. The parameters of TSPred_continuous_onGPUIdx() are as follows:

3.2.1 Parameters of TSPred_continuous_onGPUIdx():

 'groupQuery_vec' -- a set of query vector for prediction, it can be multiple queries for mulitple sensors,
	the relations between sensors and queries are defined by a vector groupQuery_blade_map as show below
 'contPrdStep' -- the step for making continuous prediction
 'groupQuery_blade_map' -- map the query to corresponding data blade (one data blade is just the time series of one sensor)
 'y_offset' -- the h-step ahead prediction for time series prediction.
                       for one step ahead prediction,y_offset==0
                       for multiple steps ahead prediction, y_offset = mul_step - 1
 'pred_selector' -- to select the prediction model
 			pred_selector = 1, smiler-AR
 			pred_selector =0, smiler-GP (refer to section 5.2 of the paper SMiLer)
 'weightedTrn' -- whether to do weighted training process, i.e. give different weight of the kNN results based DTW distance
 'selfCorr' -- whether to use auto-tuning mechanism with self-adaptive prediction to determine the weight of K and L

4. Dependencies

The current code depends on the following external libraries:

Armadillo (http://arma.sourceforge.net/), is a high quality C++ linear algebra library.

OpenBLAS (optional, http://www.openblas.net/), is an optimized BLAS library. OpenBLAS is not mandatory, but it helps Armadillo achieve better performance. (You also can use BLAS, ATLAS and LAPACK together to replace the OpenBLAS.)

Note: The project is created by Nsight IDE.

xinchoubiology/Smiler