eco routing and eco driving the reduction project€¦ · 2012-11-15 · josif grabocka, umer...

REDUCTION

Eco Routing and Eco Driving— The REDUCTION Project

Josif Grabocka Umer Khan Lars Schmidt-Thieme

Information Systems and Machine Learning Lab (ISMLL),University of Hildesheim, Germany

Arbeitskreis Informationstechnologie, Delphi, Bad Salzdetfurth,November 15, 2012

Josif Grabocka, Umer Khan, Lars Schmidt-Thieme, ISMLL, University of Hildesheim, Germany

Arbeitskreis Informationstechnologie, Delphi, Bad Salzdetfurth, November 15, 2012 0 / 31

REDUCTION

Outline

1. The REDUCTION Project

2. Eco Routing I: Estimating Fuel Consumption from GPS Data

3. Eco Routing II: Contextualized Estimation of Fuel Consumption

4. Eco Driving: Acceleration/Deacceleration Recommendations

5. Distributed Data Mining in Vehicular Networks

6. About ISMLL

7. Conclusion



REDUCTION 1. The REDUCTION Project

Outline






6. About ISMLL

7. Conclusion




REDUCTION Overview




REDUCTION Vision

I Enable fleet managers in EU to “go green” by using advanced ICTsolutions for substantially reducing mileage and fuel consumption.Based on historic and real-time data about driving behaviour, routinginformation and emission measurements.




REDUCTION Partners

I Project Duration: 01.09.2011 – 31.08.2014

I Project Homepage: http://reduction-project.eu




REDUCTION Field TrialsI. flex routes (taxis): FlexDenmark II. fixed routes (busses): CTL,

Nikosia

III. Multi-modality (trains & taxis): TrainOSE



REDUCTION 2. Eco Routing I: Estimating Fuel Consumption from GPS Data

Outline






6. About ISMLL

7. Conclusion




Data-Source Types

High frequent GPS: Trajectory-to-Eco

Work on Eco Routing I has been carried out by University of Aalborg, Universityof Aarhus and FlexDenmark.




Framework of EcoMark




Spatial Network Lifting

I MotivationI Eco-routes are highly correlated to the elevation of road segments.I TomTom states ITS using 3D spatial network will result in an annual fuel

saving of 6 Billion USD in American. [L. Sugarbaker et. al. 2012]

I 3D spatial network is constructed by using a laser scan point cloud for lifting

a 2D spatial networkI Yields a model with high accuracy




Comparison and Analysis




Comparison: Aggregated Models - 1




Comparison: Aggregated Models - 2



REDUCTION 3. Eco Routing II: Contextualized Estimation of Fuel Consumption

Outline






6. About ISMLL

7. Conclusion




Eco-Routing

I Routing provides the fastest route from a start point to a destination,i.e., the one with the shortest travel time.

I Eco Routing provides the most fuel efficient route from a start point toa destination,i.e., the one with the lowest fuel consumption.

I Routing requires Speed/Travel Time Maps that provide estimatedtravel times for each road segment.

I Eco Routing requires Fuel/Eco Maps that provide estimated fuelconsumption for each road segment.

I Travel times and fuel consumption can be estimatedI from heuristics: taking into account length of the road segment,

elevations, speed limits etc.I from observed data: take the average.




Contextualized Eco-RoutingI More accurate estimations can be made when taking context into

account such asI vehicle type,I driver,I time of day,I weather,I road conditions,I traffic conditions, etc.

I from observed data, taking averages:I build a map for each context value,

e.g., a map each for mornings, afternoons and nights.I problem: observations in each cell become sparse,

esp. when combining several contexts.

REDUCTION solution:

estimate travel time and fuel consumption for a road segmentand some context using a probabilistic model.




Example (1/3)I 3 road segments: S1,S2,S3.I 3 vehicles: V 1,V 2,V 3 (context).

Figure: Driver 1 travels across all the three segments 1, 2 and 3, while the Driver2 passes through segments 1 and 2. Finally Driver 3 has an itinerary involvingonly segment 3.




Example (2/3)

Observed travel times can be represented in a vehicle-vs-segment matrix:

T =

S1 S2 S3

V1 6 4 8V2 3 2 ?V3 ? ? 6

I How long will vehicle 2 need to pass road segment 3 ?

I How long will vehicle 3 need to pass road segment 1 or 2 ?




Example (3/3)Find a low rank / low norm matrix factorization to explain the observeddata:

T = Φ ·Ψ, T ∈ R3×3,Φ ∈ R3×1,Ψ ∈ R1×3

T =

6 4 83 2 ?? ? 6

≈ 2

11.5

· ( 3 2 4)

Use the factorization to predict unobserved cells:

T̂2,3 = Φ2,1 ∗Ψ1,3 = 4

T̂ =

6 4 83 2 44.5 3 6




Matrix Factorization

Reconstruct original matrix, X ∈ RN×M , via lower dimensionality matrices,Φ ∈ RN×K , Ψ ∈ RK×M , while K << M.

X ≈O Φ ·ΨO = { (i , j) | Xi ,j 6= ?}

Optimize the regularized loss function:

argminΦ,Ψ

L(Φ,Ψ;X , λΦ, λΨ) := ||X − Φ ·Ψ||2O + λΦ||Φ||2 + λΨ||Ψ||2

Factorization models can efficiently be learnt via Stochastic GradientDescent.




ExperimentsVery preliminary results:

I infati dataI 20 cars xI 837 road segments yI 2741 car/road segment travel time t observations (xi , yi , ti ):

t ∈ [1s, 2023s]

I split randomly 50:50 into train and test set

I trained on train set, evaluated via RMSE on test set:

RMSE(t̂, t) :=

√√√√1

n

n∑i=1

(t̂i − ti )2, t̂i := t̂(xi , yi )

method RMSE [s]

averages per road segment 167.7averages per road segment and vehicle 143.7




Predicting Fuel Consumption

I Fuel consumption can be estimated by the same type of modelsI with fuel consumption as target variable instead of travel times.

I requires per segment measurement of fuel consumptionI e.g., via CanBUS data

I several context attributes can be taken into account by moving fromMatrix Factorization to Tensor Factorization.



REDUCTION 4. Eco Driving: Acceleration/Deacceleration Recommendations

Outline






6. About ISMLL

7. Conclusion




Different Driving Behaviors

[www.toyota.com.au]




Eco Driving Aspects

I Eco Analytics:I provide insight into dependency of fuel consumption on different

contexts(driver behavior, vehicles, technical conditions, etc.)

I Local Eco RoutingI lane changeI finding parking lotsI . . .

I Acceleration/Deacceleration recommendationsI traffic light aheadI speed limit aheadI slow cars ahead etc.




Acceleration/Deacceleration RecommendationsGiven

I a planned path P (including elevations),I a current velocity v0,I a sequence A of acceleration/deacceleration actions, andI a context description C (vehicle, driver, weather, road etc.)

predict the velocity at the end of path P:

v̂(P, v0,A,C )

Such a model can be used to search for the sequence A ofacceleration/deacceleration actions that

1. leads to a target velocity vend at the end of the pathI stop, speed limit

2. is most fuel efficientI using a prediction model for fuel consumption

A special case: recommend when to coast (A :≡ 0).




Acceleration/Deacceleration RecommendationsGiven

I a planned path P (including elevations),I a current velocity v0,I a sequence A of acceleration/deacceleration actions, andI a context description C (vehicle, driver, weather, road etc.)

predict the velocity at the end of path P:

v̂(P, v0,A,C )

Such a model can be used to search for the sequence A ofacceleration/deacceleration actions that

1. leads to a target velocity vend at the end of the pathI stop, speed limit

2. is most fuel efficientI using a prediction model for fuel consumption

A special case: recommend when to coast (A :≡ 0).Josif Grabocka, Umer Khan, Lars Schmidt-Thieme, ISMLL, University of Hildesheim, Germany


REDUCTION 5. Distributed Data Mining in Vehicular Networks

Outline






6. About ISMLL

7. Conclusion




Distributed Data Mining

I Distributed predictive models, which are communication efficient,scalable, asynchronous, and robust to dynamic topology, whichachieve accuracy as close as possible to centralized ones.

Figure: DDM in Vanets




Reduced Support Vector Machinesfor Distributed Data Mining (Khan et al. [2012])

I Reduced Support Vector Machines (RSVM) for distributedvehicle-to-vehicle data mining in VANETS

I Allow only a subset of data as support vectors and solve a smalleroptimization problem

I Considering the SVM optimization problem:

minimize1

2αT (Q +

I

2C)α− eTα

subject to:

yTα = 0, and αi ≥ 0, i = 1, ..., l

I Where e is the vector of all ones, Q is an l by l positive semi-definitematrix, Qij = yiyjK (xi , xj), and K (xi , xj) = Φ(xi )

TΦ(xj)



REDUCTION 6. About ISMLL

Outline






6. About ISMLL

7. Conclusion




Research Goals

supervised prediction:learn models from data where the correct decision hasalready been observed in the past↔ unsupervised learning: detect pattern, groups etc. in data

complex data:learn models from data that cannot naturally be described asjust instances with attributes,

e.g., sequences, images, relational data, etc.

complex decisions:learn models for complex decisions that cannot naturally bedescribed as a scalar,

e.g., multi-label classification, object identification, etc.




From Application Problems to ML Problems

predictionrating

contextualized

predictionfuel consumption

database recorddeduplicaton

referencerecognition

...topicclassification

identifying productsin offers

spam filteringissue identification

...PROBEMSAPPLICATION










...PROBEMSAPPLICATION

sparse two−moderegression

predictionrating ...

contextualized


relationalclassification identification

object ...MACHINE LEARNINGPROBLEMS





?exhaustivecategorisation?






...APPLICATIONPROBEMS

sparse two−moderegression

predictionrating ...

contextualized


MACHINE LEARNINGPROBLEMS

relationalclassification identification

object ...



REDUCTION 7. Conclusion

Outline






6. About ISMLL

7. Conclusion



REDUCTION 7. Conclusion

Conclusions

I Eco Routing and Eco Driving are the fundamental building blocks foreco-friendly transport addressed in REDUCTION.

I Estimating Fuel Consumption from GPS Data can be greatlyimproved by elevation maps.

I Contextualized Estimation of Fuel Consumption can be efficientlyaccomplished by factorization models.

I Acceleration/Deacceleration recommendations can solve many ecodriving tasks and require a velocity and fuel consumption prediction.

I Distributed Data Mining in Vehicular Networks can help to bringmachine learning methods to vehicular networks.



REDUCTION

References IM. Umer Khan, Alexandros Nanopoulos, and Lars Schmidt-Thieme. Experimental evaluation of reduced support vector

machines and relevance vector machines for communication efficient distributed classification in peer-to-peer networks. Insubmitted, 2012.



eco routing and eco driving the reduction project€¦ · 2012-11-15 · josif grabocka, umer...

Documents