Está en la página 1de 16

Iterated Local Search for the Multiple Depot Vehicle

Scheduling Problem

Benot Laurent
Jin-Kao Hao
(2009)
S
avio S. Dias

August 11, 2016

Summary
Introduction
Multiple Depot Vehicle Scheduling Problem
Iterated Local Search
Computational Results
Conclusions

2 of 16

Introduction


The problem itself;

NP-hard to depot number 2;

Some strategies to consider: clustering-first & scheduling-first;

Previous metaheuristic approaches: LNS & TS (Pepin et al., 2008);

ILS with block moves neighborhood ejection-chain based.

3 of 16

Multiple Depot Vehicle Scheduling Problem


The MDVSP

Input Data:




Feasibility constraints:





Trips;
Vehicles and depots;
Transfers without passengers (pull-in, pull-out and deadhead trips);
Complete cover;
Feasible sequence;
Depot attachment;
Depot capacity;

Minimization of:



4 of 16

Fixed costs related to scheduled vehicles (primary goal);


Operational costs (secondary goal).

Multiple Depot Vehicle Scheduling Problem


Formulation

Let n, m, p be the number of trips, depots and vehicles respectively;

T = {t1 , t2 , ..., tn } the set of trips ordered by starting time;

(si , ei ) the starting and ending time of the trip ti T ;

ij , (i, j) {1, ..., n}2 , i < j the transfer time from the end of ti to
the start of tj ;

H = {n + 1, ..., n + m} the set of depots;

rh the capacity of depot h;

V = {1, ..., p} the fleet of vehicles;

Vh V the vehicles housed in depot h;

Cij , (i, j) {1, ..., n + m}2 , i < j the cost of transfers from i to j.
Note that cij = if i j or ti tj is not feasible.

5 of 16

Multiple Depot Vehicle Scheduling Problem


Solution Representation

A solution : T V , i.e. a vector of length n w/ vehicles values;

The cover, depot attachment and capacity constraints are satisfied;

To satisfy the feasible sequence: v V , (ti , tj ) T , i < j e


((ti ) = (tj ) = v ) ei + ij sj ;

Binary variable xij deduced from to represent the links between


trips and depots;

The objective function is then:


f () =

X
(i,j){1,...,n+m}2

6 of 16

cij xij

Iterated Local Search


Initial Solution

Three phases:





Reduction to SDVSP (Polynomial-time solvable) and then solve it


with Freeling et al. (2001) auction algorithm;
Using a greedy approach to assign the trips to a free vehicle with the
least cost;
Solve the SDVSP to each depot;

The first phase guarantees the minimum number of vehicles.

7 of 16

Iterated Local Search


Local Search & Neighborhood structures




A descent local search method;


Three neighborhoods:



Nshift : transfer a trip from a vehicle to another;


Nswap : exchange two vehicles of two different trips;

...and the Block-moves Neighborhood:







8 of 16

Ejection-chain based;
Exploration more extensive and effective;
Shift consecutive trips T 0 , from some vehicle v to a random one, v 0 ;
Due to constraints violations, some repairs may be needed.

Iterated Local Search


The Block-Move Neighborhood

9 of 16

Iterated Local Search


Local Search & Neighborhood structures

Intermediate technique of the RNVD;

Evaluation (and generation) of a fraction from the whole


neighborhood;

Introduction of a parameter [0, 1] to determine this fraction;

Number of iterations without improvement used as stopping


criterion.

10 of 16

Iterated Local Search


Perturbation Mechanisms & Criteria

Two proposed mechanisms:




Do moves taking into account only the increasing of the number of


used vehicles;
Reorganization of the trips in each depot;

Acceptance criterion privileges improvement (i.e. improvement of


the best by the local search);

Computational time used as stopping criterion.

11 of 16

Computational Results
Benchmarks & settings

The Pepin et al. (2008) benchmark were used;

Six categories, w/ variations: m {4, 8} and


n {500, 1000, 1500};

Five instances per category, leading to 30 test instances;

Algorithm coded in C++ and compiled w/ VC++ 8.0 (flag -O3);

The tests were carried on a 2GHz T7200 Intel Core w/ 2GB RAM
running a Windows XP;

Parameters: = .002; = .5; and = 1000;

Gap =

12 of 16

f ()f (best )
f (best )Pnbvehi (best ) ,

where P is set to 10000.

Computational Results
Comparison of Neighborhoods

13 of 16

Computational Results
Comparative Results

The ILS were compared to 4 state-of-art previous algorithms: A


column generation heuristic; a Lagrangian relaxation; a LNS; and a
TS w/ Nshift and Nswap ;

The time shown in the previous heuristics correspond to the


minimum amount of time;

A 1.33 ratio were applied to the computational time in order to


equalise the comparison;
The times set were:




14 of 16

110, 930, 3060 s for the 4 depots sets;


165, 1200, 4260 s for the 8 depots sets;

Computational Results
Comparative Results

15 of 16

Conclusions


An effective construction method;

A powerful Neighborhood (Block-moves) and RVND technique;

A good trade-off between time and good solution in comparison


with the other metaheuristics methods;
My Thoughts:




16 of 16

The stopping criterion (?);


The computational time comparison due to the ratio (?);

También podría gustarte