About me

• Geography / GIS background
• Worked 3 ys for virtualcitySYSTEMS
• Research assistant @ Beuth HS
• Founder of the MAGDa team
• ♥ using and teaching PostGIS
• Promoting FOSS4G
• @FlxKu

About Beuth HS

• Founded 1971 as Technische Fachhochschule (TFH)
• Located in wild wild Wedding
• 12.000 students
• Widest range of degree courses in applied engineering in Berlin and Brandenburg
• Hosted the german FOSS4G in 2014

Arbeitskreis Data Science

• Joint group of Profs from Math, Machine Learning, Computer Vision, Databases, Text Mining, Geotemporal Data Mining
• Regular PhD Colloquiums and Journal Clubs

             Project

• Funded by German government and DLR
• Goal: Build up a mobility service plattform
• Addressing small companies
• Crowdsource traffic related data
• Explore / extend spatial data mining tools

               Mobility Stack

               Mobility Stack

• Processing sensor streams
• Calculate travel times
• Predict traffic states
• Route optimization
• Smart scheduling

Problem definition

• Many ITS do only traffic monitoring
• Rule-based systems reacting to current traffic
• Great uncertainty in traffic messages

Travel time to incident in min

Pape, S. & Körner M. (2016). Verkehrslageprognose unter Berücksichtigung der dynamischen Kapazitäten an LSA-abhängigen Knotenpunkten zur qualitativen Aufwertung der Verkehrslageinformation im Verkehrsmanagementsystem VAMOS. 25. Verkehrswissenschaftliche Tage 16. und 17. März 2016 in Dresden. (in german)

Tasks

• Predict future measurements at static sensors
• Predict travel times for trajectories
• Predict anomaly patterns (e.g. congestion)

Relation between macroscopic measures

Speed

Flow (Occupancy)

Density

Treiber, & Kesting (2013): Traffic Flow Dynamics - Data, Models and Simulation. Springer. 26.

Theory

Free Flow vs. Congestion

https://www.civil.iitb.ac.in/tvm/1111_nptel/512_FundRel/plain/plain.html

Speed-Flow Diagram

http://www.cl.cam.ac.uk/~rg31/research/

Temporal Dependency

Waterloo, S. (2017): Analyse und Bereinigung unvollständiger und fehlerhafter Messreihen von Verkehrssensoren. Betreute Bachelor-Thesis.

Temporal Dependency

Waterloo, S. (2017): Analyse und Bereinigung unvollständiger und fehlerhafter Messreihen von Verkehrssensoren. Betreute Bachelor-Thesis.

Spatial Dependency

Pfeifer, P.E. & Deutsch, S.J. (1981): Seaosonal Space-Time ARIMA Modeling

Adjacency Matrix

How to do it?

• Exponential smooting, moving average
• Autoregressive time series models (e.g. ARIMA)
• kNN search on historic data
• Baysian Networks
• Support Vector Machines (SVM)
• Artificial Neural Networks (ANN)

What works best?

• Hard to say. Papers usually favor one approach.
• How important are non-linear patterns?
• How is the performance & scalability?
• Does it work for a whole city?
• Do we need a combination of different methods?

So what is Deep Learning?

• It's Machine Learning with neural networks
• "Deep" reflects the anatomy of these nets
• Mostly supervised learning
• Ideal for fuzzy data (images, audio, text)
• Outperfoms other approaches
• New easy-to-use frameworks for DL

Yes, it's a hype

#googletrends

#GWF2017

#alphago

Images

#deeposm

#terrapattern

#orbitalinsight

Convolutional Nets

#neuralnetworkzoo

(c) Fjodor Van Veen, Asimov Institute

Input Cell

Kernel

Convolutional

Hidden Cell

Output Cell

Memory Networks

#colah_on_LSTMs

=

The entire net repeated and interconnected

Sequences?

Single values (many-to-one)

Sequences (many-to-many)

Location?

Test Data

• ~ 130 inductive loops (speed, occupancy etc.)
• 9 years of data (atm testing only 1 month of 2014)
• 80% training, 20% test

Time Series

Workflow

• Take moving average from each sensors as input ...
• ... to predict a future value for one sensor
• 5, 10, 15, 30, 45 minutes offset to predict
• Narrow down to neighbours (sparse input matrix)
• Feed input vectors into network randomly

Isochrones

Test Setup

• Network: Feed Forward
• Layers: 1 input, 1 hidden, 1 output
• Activation Function: Sigmoid
• Optimization: Stochastic Gradient Descent
• Loss Function: RMSE
• Other Hyperparams:
  
  • n Neurons = 60 (sensors)
    
  • Learning rate = 0.01
    
  • Batch Size = 20
    
  • Iterations = 10000
    

Feed Forward NN Inputs

• FFNNsingle fed only with target sensor
• FFNNNN fed only with neihbours of target sensor
• FFNNNN+ fed with neighbours incl. target itself
• FFNNall fed with values from all sensors

FFNN mimicing RNNs

• Included sequence information into input by appending the input matrix (idea from Polsen et al. 2015)
• mFFNNsingle fed only with target sensor
• mFFNNNN fed only with neihbours of target sensor
• mFFNNNN+ fed with neighbours incl. target itself
• mFFNNall fed with values from all sensors

Results

Results

Results

MA = 15 Min; Sequence Size = 10 time steps

Results

MA = 15 Min; Sequence Size = 5 time steps

Not there yet ...

• Results are a little "too good". Why is that?
• MA is probably too big. Prediction task is too easy.
• What can we really say about spatio-temporal patterns?
• Why do neighbours not have a positive effect?
• How good is the prediction when anomalies occur?

Roadmap 2017

• Present existing results and idea DONE
• Create R package to repair missing data DONE
• How good are our results really?! WIP
• Use LSTM NN to learn sequences DONE
• Use trajectories instead of sensors TODO
• Let TU Dresden test it with realtime data TODO
• Create a web-service for ExCELL WIP
• Concept for continous learning TODO
• Compare Deep Learning Frameworks DONE
• Compare with STARIMA WIP
• Compare with SVM WIP