A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

15
A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks */20 A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks Universidad de la República, Montevideo, Uruguay Universidad de Málaga, Málaga, Spain

Transcript of A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

Page 1: A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks */20

A Multiobjective Evolutionary Algorithm forInfrastructure Location in Vehicular Networks

Universidad de la República, Montevideo, Uruguay Universidad de Málaga, Málaga, Spain

Page 2: A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

Index

1. Introduction2. The RSU Deployment Problem3. A Multiobjective Evolutionary Algorithm for RSU-DP4. Experimental Analysis 5. Conclusions and Future Work

Page 3: A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

Introduction

• Vehicular Ad-Hoc Network (VANET): vehicles + roadside units (RSU)

• RSU deployment problem: find number, type, and location of RSUs to maximize QoS, while satisfying deployment cost requirements

This is a hard-to-solve optimization problem on city-scaled areas. Heuristics and metaheuristics allow computing good infrastructure

designs in reduced execution times.

Page 4: A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

The RSU Deployment Problem (RSU-DP)Definition• A set of road segments S = {s1 , . . . , sn}. Each segment si is defined

by a pair of points pj , pk P = {p∈ 1 , . . . , pm} identified by (lat, long).RSUs can be placed at any location within each segment si

• An estimation of the number of vehicles per time period across each segment NV(si)

• The average vehicle speed for each segment sp(si).• A set of RSU types T = {t1 , . . . , tk}, each one with a given cost and

coverage areaFind a set of locations and the type of each RSU to deploy, to maximize the number of vehicles served and

simultaneously minimize the total cost of deployment

Page 5: A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

The RSU Deployment Problem (RSU-DP)Related work• Lochert et al. (2008): GA and VANET simulator to evaluate the QoS

– applied to a given area of Brunswick, Germany, considering 100 possible predefined locations for RSUs

– good cost/utility trade-off using between 10 and 30 RSUs• Cavalcante et al. (2012): compared GA against a greedy approach

for the maximum coverage with time threshold problem– used data from four Swiss regions– the GA outperformed the greedy approach

No related works using explicit multi-objective methods. We propose using the NSGA-II algorithm to solve the RSU-DP.

Page 6: A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

MOEA for the RSU-DPSolution encoding• Solutions are represented as real arrays of length n = #S.• RSU type: integer part of the real number (0 stands for no RSU,

1...k represent types t1… tk )• Position within the segment: fractional part of the real number,

maps the interval [0, 1) to points in segment [pj , pi).

Page 7: A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

MOEA for the RSU-DPEvolutionary operators• Population is randomly initialized using reals from [0, k+r],

with k = #T and r [0, 1)∈• Intermediate Recombination; offspring of parents x and y satisfy

αixi + (1 − αi )yi and βiyi + (1 − βi)xi, with αi , βi randomly chosen in [−p, 1 + p] for a given p.

• Ad-hoc mutation operator:– with probability πA removes the RSU (if any) from the segment– with probability πB changes the type of the RSU (if any) to a

random type picked uniformly in T– with probability 1 − πA − πB applies a Gaussian Mutation with

standard deviation σ.

Page 8: A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

MOEA for the RSU-DPComputing the objective functions• Total cost: adds the cost (according to the type) of each RSU placed• QoS:

– Find the portions of segments covered by the RSUs– QoS considers the amount of vehicles and speed in each

portion covered.

Page 9: A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

Experimental analysisInstances, configuration, and environment• Real map of Málaga (Spain) covering 42.557 km2: 121 segments with

lengths between 55 and 1556 meters• Real traffic data: normal, low and high traffic scenarios• Three types of IEEE 802.11p omni-directional antennas:

• Parameter configuration: population size=72, crossover prob.=0.95, mutation prob.=0.01, πA=0.5, πB=0.25, σ=0.25

• Parallel evaluation on AMD Opteron 6172 24 cores 2.10 GHz, 24GB RAM [Cluster FING, HPC facility at Universidad de la República]

Page 10: A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

Experimental analysisComparison against two greedy strategies• Greedy QoS

– Places RSUs sequentially over non-covered segments at a random position

– Starts with those with higher ratio between number of vehicles and average speed

– A segment is considered covered if it has a portion of λ inside the coverage area of any RSU

• Greedy cost – analogous, but stops when the QoS is equal to α·Q – Q: best QoS value achieved by the greedy QoS with λ = 0.75– α [0, 1]∈– Experimental analysis: λ {0.9, 0.95, 1.0}, α {0.7, 0.75, 0.8}∈ ∈

Page 11: A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

• Up to 73.9% improvement in QoS with same cost than greedy cost • Up to 365.1% improvement in cost with same QoS than greedy QoS

• Small generational distance values and RHV suggest good distributions and convergence to an ideal Pareto front

Experimental analysisNumerical results

Page 12: A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

Experimental analysisPareto front (PF), normal traffic• Global PF vs. best greedy solutions, normal traffic scenario

Page 13: A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

Conclusions

• Main contribution: multiobjective evolutionary approach to solve the problem of locating roadside infrastructure for vehicular networks

• The experimental analysis was performed using real traffic information from the city of Málaga.

• The proposed MOEA has shown good problem solving capabilities:– computing accurate Pareto fronts (according to standard

multiobjective metrics)– significantly improving two greedy heuristics (up to 365.1% in

cost and 73.9% in QoS)

Page 14: A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

Current and future work

• Overlapping antennas need to be considered in the model.

Page 15: A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

A Multiobjective Evolutionary Algorithm for Infrastructure Location in Vehicular Networks

Current and future work

• Considering different QoS metrics including:– additional information, such as accidents– different media transmitted by the RSUs (text, audio, video).

• Comparing against other techniques:– Different greedy methods

• PageRank algorithm• Knapsack

• Extending the experimental analysis to other cities (difficulty in getting access to data).