Part of the Joint Project by HKBU, HKIVE and Several Local Mobile Service Providers for Accurate Low-cost Mobile localization Support Vector Regression for Location Estimation Using GSM Propagation Data Dr. Chun-hung Li Department of Computer Science Hong Kong Baptist University June, 2003

GSM Localization via Missing Value Insensitive Support Vector Regression Contents Introduction Related Works SVR via Missing Value Insensitive Kernel Simulation & Field Test Q & A

GSM Localization via Missing Value Insensitive Support Vector Regression Introduction Task To estimate the location of a mobile device using the information based on the GSM Networks Approach -- Network-based Solutions Provide the location service using the network information without modifying the mobile phone Baseline Accuracy Federal Communications Commission rule - 100m (67% of the time)

GSM Localization via Missing Value Insensitive Support Vector Regression Introduction – GSM Network Information Returned from the mobile phone side 1. Serving Cell ID 2. BSIC 3. BCCH No 4. Received signal strength (dBm) Other Station Information Station Position (x & y) Height Bearing Cell Type Antenna Type Station Power strength (dBm) …… 1 324

GSM Localization via Missing Value Insensitive Support Vector Regression Related Works - Network-based solution Precise time and direction based methods - TOA: Time of Arrival - AOA: Angle of Arrival - TDOA: Time-Difference of Arrival - Require Synchronization Clock or Smart Antennas Signal Strength Attenuation Modeling Approach - Mapping signal strength into distance -- e.g. Free Space Model, HATA model, … - Recover coordinate from distance -- Cell-ID, Weighted CG -- Tri-lateration

GSM Localization via Missing Value Insensitive Support Vector Regression Related Works – Weighted CG & Cell-ID Based on Free Space Model – The distance and the received signal strength is an inversely proportional function – Or Approximation: Weighted Central of Gravity (CG) –Smaller Distances -> nearer to stations –If N is 1, obtain the Cell-ID Method where N is the number of neighboring base stations, Δs is the signal strength falloff in dBm

Related Works – Circular Trilateration Transmitter Estimated mobile location r1 r2 r3 GSM Localization via Missing Value Insensitive Support Vector Regression

Related Works – Machine Learning Approach More robust calibration of Propagation Models Statistical Modeling Approach Directly map signal strength to location output Wireless LAN Positioning via Neural Network, Support Vector Classification/Regression Fingerprinting Method GSM Localization via Missing Value Insensitive Support Vector Regression

Why using Machine Learning Approaches Hard to Obtain a Parametric Model Terrain Factors, multi-path, occlusion, … Noise Measurement, Weather Condition, … Comparably Easy to get a lot of data Fit a nonparametric model to the data No need for domain experts/domain models Changes in models/parameters can be re-learned GSM Localization via Missing Value Insensitive Support Vector Regression

Adopting a mapping to transform all signal strength readings at a location into a series of descriptors: E.g. Linearly regress the series of descriptors into the position output Introduction to Support Vector Regression GSM Localization via Missing Value Insensitive Support Vector Regression W is of the same length as the long descriptor vector

w by solution is the linear combination of a set of descriptor vectors from l training data E.g. Location output (x or y) : The key is to seek a Kernel function Introduction to Support Vector Regression – Cont. GSM Localization via Missing Value Insensitive Support Vector Regression Where r (i) denotes the i-th signal vector used for training

e.g. RBF Kernel: S is a severely sparse vector Only 3~9 signals are retrievable e.g. two sample signal reading Vectors: Impute empty cells by values: Too many! & What’s the physical meaning? Incompetent Conventional Kernels GSM Localization via Missing Value Insensitive Support Vector Regression Station r-71-60N s N-72N

Sum of Exponential Kernel (SoE) Where It is a valid kernel by proof Recently proved to be a variant of the 1st-order RBF-ANOVA Kernel A New Missing Value Insensitive Kernel GSM Localization via Missing Value Insensitive Support Vector Regression

A Kernel Matrix Evaluated from SoE GSM Localization via Missing Value Insensitive Support Vector Regression

Experimental Results – Simulation Study GSM Localization via Missing Value Insensitive Support Vector Regression Model adapted from [Roos 2001] Adding Occlusion and Noise effects Experiment Settings 30 km 2 Data Collection Region 640 Training Markers, 200 Testing Markers 64 Base Stations, 8 receivable Roos RBF without any missing value handling SoE Mean Error (m)

Data Collection GSM Localization via Missing Value Insensitive Support Vector Regression Experimental Results – Field Data Test

GSM Localization via Missing Value Insensitive Support Vector Regression Experiment Settings A 350 x 550m data Collection Region Total 15 Markers 120 set of readings / marker 50 Base Stations, 7~9 receivable CGCT mean Error(m)

Experimental Results – Field Data Test GSM Localization via Missing Value Insensitive Support Vector Regression Experiment Results For SVR Training: 9 Markers for Training Multiple sets of readings from each training marker For SVR Testing: 1.Predict one location for a single set of readings 2.Predict one location for multiple sets of readings acquired at the same site and in a short interval

1) 8 of 120 sets of training readings from each of the 9 of 15 markers 2) 120 sets of testing readings from the remain 6 of 15 markers 3) mean error = 47m GSM Localization via Missing Value Insensitive Support Vector Regression

1) predict 120 sets of readings in each testing marker to one location 2) interval: 2 min 3) mean error = 21m GSM Localization via Missing Value Insensitive Support Vector Regression

or shown in following diagram: GSM Localization via Missing Value Insensitive Support Vector Regression

Q & A GSM Localization via Missing Value Insensitive Support Vector Regression