arXiv:1502.03699v1 [cs.NE] 12 Feb 2015
1
Analysis of Solution Quality of a Multiobjective
Optimization-based Evolutionary Algorithm for
Knapsack Problem
Jun He, Yong Wang and Yuren Zhou
Abstract
Multi-objective optimisation is regarded as one of the most promising ways for dealing with constrained optimisation problems
in evolutionary optimisation. This paper presents a theoretical investigation of a multi-objective optimisation evolutionary algorithm
for solving the 0-1 knapsack problem. Two initialisation methods are considered in the algorithm: local search initialisation and
greedy search initialisation. Then the solution quality of the algorithm is analysed in terms of the approximation ratio.
I. INTRODUCTION
Consider the problem of maximizing an objective function,
max
~x
f(~x), subject to g(x) ≤ 0. (1)
The above constrained optimisation problem can be transferred into an unconstrained bi-objective optimisation problem. That
is to optimize the original objective function plus to minimize the constraint violation simultaneously:
max
~x
f(~x),
min
~x
v(~x),
(2)
where v(~x) is the degree of constraint violation, given by
v(~x) =
0, if g(~x) ≤ 0,
g(~x), otherwsie.
(3)
The use of multi-objectives for single-objective optimisation problems could be traced back to 1990s [1]. This methodology
has been termed multiobjectivisation [2]. Using multiobjectivization sometimes may help the search more efficient as shown
in [3]–[6].
According to the survey [7], multi-objective optimisation is regarded as one of the most promising ways for dealing with
constrained optimisation problems in evolutionary optimisation. A constrained optimisation problem is often transformed into
a bi-objective optimisation problem, in which the first objective is the original objective function and the second objective is
the degree of constraint violation [8]–[11]. After this transformation, Pareto dominance is frequently employed to compare
individuals. Currently the research in this area is very active [7]. For example, a self-adaptive selection method is proposed
recently in [12], which aims to exploit both non-dominated solutions with low constraint violations and feasible solutions with
low objective function values. Multi-objective optimisation is combined with differential evolution in [13] and an infeasible
solution replacement mechanism is proposed. A dynamic hybrid framework is presented in [14], where the global and local
search models are implemented dynamically according to the feasibility proportion of the population.
This paper aims at analysing the solution quality of evolutionary algorithms (EAs) in terms of the approximation ratio. It
is not intended to demonstrate that EAs are able to compete with problem-specific approximation algorithms, since this is
unlikely in most cases. Nevertheless, it is still necessary and important to understand the solution quality of EAs, so the EAs
with arbitrarily bad solution quality could be avoided in applications. The analysis of the approximation performance of EAs
has attracted a lot of interests in recent years [15]–[17].
This paper investigate an existing multiobjective optimization-based EA [9] (MOEA) for solving constrained optimisation
problems. The MOEA originally is designed for continuous optimization. Here it is adapted for solving the 0-1 knapsack
problem. Although experiment results show its performance is good, no theoretical analysis exists for this MOEA [9]. This
motivates our rigorous analysis.
The remainder of the paper is organized as follows. The 0-1 knapsack problem and approximation ratio are introduced in
Section II. The MOEA with the local search initialisation is analysed in Section III . Section IV is devoted to the analysis of
the MOEA with the greedy search initialisation. Section V concludes the article.
This work was partially supported by EPSRC under Grant No. EP/I009809/1 (He), by NSFC under Grant Nos. 61170081, 61472143 (Zhou), 61273314
and by the Program for New Century Excellent Talents in University under Grant NCET-13-0596 (Wang).
Jun He is with Department of Computer Science, Aberystwyth University, Aberystwyth, UK
Yong Wang is with School of Information Science and Engineering, Central South University, Changsha 410083, China
Yuren Zhou is with School of Advanced Computing, Sun Yat-sen University, Guangzhou, 510006, China