SyNS'13 Session 4: Measure it!

Inaccurate Spectrum Databases?
Tan Zhang, Suman

TV whitespaces are availbe, how do we use them?

Things people are already doing:
1. Rutal internet connection
2. Smart power-grid controller
3. Bus internet

But it shouldn't interfere primary TV users, thus need to determining whitespaces.

Wardriving on Public Vehicles (spectum sensor depoyed on Metro buses)

Accuracy of Commercial Databases:
1. Good for protecting used spaces
2. But too conservative thus waste while spaces.

V-scope: opportunistic
Wireless Gateway upload data to V-Scope Servers, then Serve uses internet to distribute information to participating databases.

Callenge in measuing Primary Signal:
Weak signals could be overwhalemed b strong signals ---> Zoom-in Peak detection (narrowing capture bandwidth to reduce noise floor).
Power ---> peak based power estimation.
Model refinement --> improve any propagation models by fitting its parameters with measurements. ---> weighted regression fitting to allow good fitting even for weak signals.


Enhancing the Quality of Electric Load Forecasting Methods
Scott ALfeld, Paul Barford

Predicting Electricity: compnents, buildings, cities, countries, days, years.

This work looks at buildings and cities, not single homes.

Problem:
Many, many buildings make up cities
Easily available: Historical electricity usage
Less so: Everything else (what if you have a new building, a new event?)

Data: hourly electriy usage for 2-3 years, NAICS Code ("business type")
Task: Predict tomorrow's (hourly) load

Solution:
Latent Feature Extraction
Anomaly Detection
Priority Encoding

Latent Feature Exatrction:
Solid, but unknown features: number of floors, surface area
Fuzzy and unknown features: sensitivy to humidity, behavior during holidays, ??? (something we didn't thought of)

Anomaly Detection:
Individual Anomalies (for one building): boiler upgrade, 50%-off sale
Shared Anomalies (for multiple buildings): three-day weekends, snowstorms

Priority Encoding:
Accuracy may not be the most important thing. Peak is much more important than total usage.
Peak prediction and Discontinuous penalties: MSE: 1/n* sigma(ti - yi)^2 doesn't work because peak error are more important for your prediction than errors which are made for free hours; you also don't want false positive peaks.
Challenges: need to aligin the peak with actual peak, but not a straight line at peak value! It needs to be easy to optimize on.
We don't have an solution for that yet....


Observing Home Wireless Experience through WIFI APs
Ashish Patro, Suman

A measurementinfrastrure which could capture the "wireless experience" in home WLANs.
Questions asked:
1. What is the wireless performance at any given moment?
2. How often does poor performance occur?
3. What are the causes of poor performance?

Wise infrastructure (wireless infrastructure of in-line sensing)
Data collected:
Signal strength of devices, air timeutilization, latencies. wifi packet statistics.
Time correlation analysis
Characterizing wireless performance:
   metric: only passive and corse local measurements, capture impact of links, provide app anoglistics
    Input: air time utilization, total contention, ????

Results:
Cause of poor performance
in building 1, when performance is poor, 60% of time we have high air time utilization and high packet loss, thus indicate channel congestion
in building 2, when performance is pool, latency is high and packet loss is high, indication a week signal source

Interference of Microwave oven:
Interference is short, but present in most buildings