1
BEARS SinBerBEST APEC: Auto Planner for Efficient Configuration of Indoor Positioning Systems Ming Jin, Ruoxi Jia, Costas Spanos # routers # fingerprints WiFi, GSM, iBeacon… • Fingerprin(ng: RADAR, Horus • Model based: RF prop. model Indoor posi(oning system Setup Cost Accuracy User preferences U(lity LocaCon Priority VisiCng Frequency Maximize! Budget APEC: design the fingerprintsbased IPS that accounts for user preferences and budget constraints IntroducCon and ContribuCon Hierarchical Bayesian Signal Model (HBSM) U(lity Op(miza(on Implementa(on of APEC Objec(ve func(on Algorithm Design • Learningtolearn to improve RSSI es(ma(on • Objec(ve as a theore(cal solu(on close to actual loss • Computa(onal tractable as a guidance for field usage Local priority map Local frequency map MEDIUM Priority LOW Freq. HIGH Frequency MED Frequency HIGH Priority LOW Priority Objec(ve: customer behavior analysis in stores User Preferences Priority weighted misclassifica3on loss 1 2 Expecta3on over visi3ng freq. 3 HBSM randomness 4 Number/loca3on of routers/fingerprints OpCmizaCon Framework Minimiza'on of the expected loss: Weighted cost of loca'on confusion Frequency Priority Misclassifica'on rate APEC Algorithm θ rt (router locations): Given M possible locations, choose K to place the routers. θ fp (fingerprints): Distribute the number of fingerprints to be collected. Tasks: Exhaustive: search all possible router/ fingerprints combinations (intractable!) Greedy: stochastically optimize through coordinate descent (heuristic!) Strategies: Hierarchical Bayesian Signal Model (HBSM) How can we make efficient use of fingerprints? PassLoss Model Fingerprints Parameter Es(ma(on Fingerprints Es(ma(on Gaussian Process Map *Neighborhood covariance func(on Top layer: Hyperpriors Mean of RSSI follows LossPath Model and Gaussian Process BoWom layer: ObservaCons Inference Es(ma(on RSSI observa(on at loca(on i for K routers that work independently HBSM: SpaCal variance Measurement error Radio map reconstrucCon: empirical Bayes and Gaussian process regression. Length (m) Width (m) Toy Case Study Heuris(c 1: increment the number of fps at many random batches of locaCons and choose the best Router set index Expected loss Op#mal (Exhaus#ve) Decrease of batch size Random selec3on Sorted from the highest loss to the lowest for each router configura5on BEST! Router set index Expected loss Top 10 router setups Random selec3on Monotone increasing Sorted from the lowest loss to the highest for each router configura5on 4 Heuris(c 2: router locaCons can be chosen assuming uniform fingerprints allocaCon Priority and Freq. Map • HIGH Priority for cubicles: automatic climate control • MED Priority for shared spaces: energy apportionment • LOW Priority for corridors Field Deployment Hypothesis 1: the expected cost is a good indicator of the actual cost of the system • Strong correlation between expected & actual cost • Predict IPS performance based on the router- fingerprints configuration Expected cost op+mized by APEC Actual cost Expected cost: Actual cost: Theore&cal Actual misclass. Hypothesis 2: APEC Greedy performs well for the actual cost of the system (solu(on superiority) Alignment of expected cots to actual cost Exp A: 5 routers Exp B: 7 routers Size represents density of fingerprints High priority areas 1 3 2 Lower priority areas Closer to routers Visualiza(on Conclusion and Future APEC: systemaCcally opCmizes the locaCons of APs and fingerprints Implement and visualize APEC configuraCon on mobile pla\orms Publica(on: APEC: Auto Planner for Efficient ConfiguraCon of Indoor PosiConing Systems, 9th InternaConal Conference on Mobile Ubiquitous CompuCng, Systems, Services and Technologies, 2015
