Chapter
Chapter
(1)
where
1, if constraint is violated
0, if constraint is satisfied
i
i
i
i
'
f
p
(x) is the penalized objective function, f(x) is the unpenalized objective function, and C
i
is a
constant imposed for violation of constraint i. This penalty function is based only on the number
6
of constraints violated, and is generally inferior to the second approach based on some distance
metric from the feasible region (Goldberg 1989, Richardson et al. 1989).
More common and more effective is to penalize according of distance to feasibility, or the
cost to completion as termed by Richardson et al. (1989). This was done crudely in the
constant penalty functions of the preceding paragraph by assuming distance can be stated solely
by number of constraints violated. A more sophisticated and more effective penalty includes a
distance metric for each constraint, and adds a penalty which becomes more severe with distance
from feasibility. Complicating this approach is the assumption that the distance metric chosen
appropriately provides information concerning the nearness of the solution to feasibility, and the
further implicit assumption that this nearness to feasibility is relevant in the same magnitude to
the fitness of the solution. Distance metrics can be continuous (see for example, Juliff 1993) or
discrete (see for example, Patton et al. 1995), and could be linear or non-linear (see for example,
Le Riche et al. 1995).
A general formulation is as follows for a minimization problem:
f f Cd
p i i
i
m
( ) ( ) x x +
1
(2)
where d
g i q
h i q m
i
i i
i
'
( ),
( ) ,
x
x
for = 1, ...,
for = +1, ...,
d
i
is the distance metric of constraint i applied to solution x and is user defined exponent, with
values of of 1 or 2 often used. Selection of C
i
is more difficult. The advice from Richardson
et al. (1989) is to base C
i
on the expected or maximum cost to repair the solution (i.e. alter the
solution so it is feasible). For most problems, however, it is not possible to determine C
i
using
this rationale. Instead, it must be estimated based on the relative scaling of the distance metrics
of multiple constraints, the difficulty of satisfying a constraint, and the seriousness of a
constraint violation, or be determined experimentally.
Many researchers in evolutionary computation have explored variations of distance based
static penalty functions (e.g., Baeck and Khuri 1994, Goldberg 1989, Huang et al. 1994, Olsen
7
1994, Richardson et al. 1989). One example (Thangiah 1995) uses a linear combination of three
constant distance based penalties for the three constraints of the vehicle routing with time
windows problem. Another novel example is from (Le Riche et al. 1995) where two separate
distance based penalty functions are used for each constraint in two GA segregated
subpopulations. This double penalty somewhat improved robustness to penalty function
parameters since the feasible optimum is approached with both a severe and a lenient penalty.
Homaifar et al. (1994) developed a unique static penalty function with multiple violation levels
established for each constraint. Each interval is defined by the relative degree of constraint
violation. For each interval l, a unique constant, C
il
is then used as a penalty function
coefficient.
C 5.2.5 Dynamic Penalty Functions
The primary deficiency with static penalty functions is the inability of the user to determine
criteria for the C
i
coefficients. Also, there are conflicting objectives involved with allowing
exploration of the infeasible region, yet still requiring that the final solution be feasible. A
variation of distance based penalty functions, which alleviates much of these difficulties, is to
incorporate a dynamic aspect which (generally) increases the severity of the penalty for a given
distance as the search progresses. This has the property of allowing highly infeasible solutions
early in the search, while continually increasing the penalty imposed to eventually move the
final solution to the feasible region. A general form of a distance based penalty method
incorporating a dynamic aspect based on length of search, t, is as follows for a minimization
problem:
f t f s t d
p i i
i
m
( , ) ( ) ( ) x x +
1
(3)
where s
i
(t) is a monotonically non-decreasing in value with t. Metrics for t include number of
generations or the number of solutions searched. Recent uses of this approach include Joines
and Houck (1994) for continuous function optimization and Olsen (1994), who compares several
penalty functions, all of which consider distance, but some also consider evolution time. A
8
common concern of these dynamic penalty formulations is that they result in feasible solutions at
the end of evolution. If s
i
(t) is too lenient, final infeasible solutions may result, and if s
i
(t) is too
severe, the search may converge to non-optimal feasible solutions. Therefore, these penalty
functions typically require problem specific tuning to perform well. One explicit example of
s
i
(t) is as follows, from Joines and Houck (1994),
s
i
(t) = (C
i
t)
1
and
2
(
1
>
2
>1), to adaptively update the penalty function multiplier, and the evaluation of
the feasibility of the best solution over successive intervals of N
f
generations. As the search
progresses, every N
f
generations the penalty function multiplier is updated based on whether the
best solution was feasible during that interval. Specifically, the penalty function is as follows,
f k f d
p k i
i
m
( , ) ( ) x x +
1
(4)
9
k
k f
k f
k
N
N
+
'
1
1
2
,
,
if previous generations have infeasible best solution
, if previous generations have feasible best solution
otherwise
Smith and Tate (1993) and Tate and Smith (1995B) used both search length and constraint
severity feedback in their penalty function which was enhanced by the work of Coit et al.
(1995). This penalty function involves the estimation of a near-feasible threshold (NFT) for each
constraint. Conceptually, the NFT is the threshold distance from the feasible region at which the
user would consider the search as getting warm. The penalty function encourages the
evolutionary algorithm to explore within the feasible region and the NFT-neighborhood of the
feasible region, and discourage search beyond that threshold. This formulation is below:
f t f F t F t
d
NFT
p feas all
i
i i
m
( , ) ( ) ( ( ) ( )) x x +
_
,
1
(5)
where F
all
(t) denotes the unpenalized value of the best solution yet found, and F
feas
(t) denotes the
value of the best feasible solution yet found. The F
all
(t) and F
feas
(t) terms serve several purposes.
First, they provide adaptive scaling of the penalty based on the results of the search. Second,
they combine with the NFT
i
term to provide a search specific and constraint specific penalty.
The general form of NFT is:
NFT
NFT
+
o
1
(6)
where NFT
o
is an upper bound for NFT. is a dynamic search parameter used to adjust NFT
based on the search history. In the simplest case, can be set to zero and a static NFT results.
can also be defined as a function of the search, for example, a function of the generation
number (t), i.e., = f(t) = t. A positive value of results in a monotonically decreasing NFT
(and thus, a larger penalty) and a larger more quickly decreases NFT as the search progresses,
incorporating both adaptive and dynamic elements.
10
If NFT is ill defined a priori, it can be set at a large value initially with a positive constant
used to iteratively guide the search to the feasible region. With problem specific information, a
more efficient search can take place by defining a tighter region or even static values of NFT.
C 5.2.7 Future Directions in Penalty Functions
Two areas requiring further research are the development of completely adaptive penalty
functions which require no user specified constants and the development of improved adaptive
operators to exploit characteristics of the search as they are found. The notion of adaptiveness is
to leverage the information gained during evolution to improve both the effectiveness and the
efficiency of the penalty function used. Another area of interest is to explore the assumption that
multiple constraints can be linearly combined to yield an appropriate penalty function. This
implicit assumption of all penalty function used in the literature assumes that constraint
violations incur independent penalties and therefore, there is no interaction between constraints.
Intuitively, this seems to be a possibly erroneous assumption, and one could make a case for a
penalty which increases more than linearly with the number of constraints violated.
11
C 5.2 References
E J Anderson and M C Ferris 1994 Genetic algorithms for combinatorial optimization: the
assembly line balancing problem ORSA Journal on Computing 6 161-173
M. Avriel 1976 Nonlinear Programming: Analysis and Methods Prentice Hall Englewood Cliffs
NJ
T Baeck and S Khuri 1994 An evolutionary heuristic for the maximum independent set problem
Proceedings of the First IEEE Conference on Evolutionary Computation 531-535
J C Bean and A B Hadj-Alouane 1992 A dual genetic algorithm for bounded integer programs
University of Michigan Technical Report 92-53 to appear in RAIRO - RO
D W Coit A E Smith and D M Tate 1995 Adaptive penalty methods for genetic optimization of
constrained combinatorial problems ORSA Journal on Computing in print
K A DeJong and W M Spears 1989 Using genetic algorithms to solve NP-complete problems
Proceedings of the Third International Conference on Genetic Algorithms 1989 124-132
H Esbensen 1995 Finding (near-)optimal Steiner trees in large graphs Proceedings of the Sixth
International Conference on Genetic Algorithms 485-490
M L Fisher 1981 The Lagrangian relaxation method for solving integer programming problems
Management Science 27 1-18
M R Garey and D S Johnson 1979 Computers and Intractability: A Guide to the Theory of NP-
Completeness W H Freeman and Co San Francisco
D E Goldberg 1989 Genetic Algorithms in Search Optimization and Machine Learning Addison-
Wesley Reading MA
A B Hadj-Alouane and J C Bean 1992 A genetic algorithm for the multiple-choice integer
program University of Michigan Technical Report 92-50 to appear in Operations Research
W-C Huang C-Y Kao and J-T Horng 1994 A genetic algorithm approach for set covering
problem Proceedings of the First IEEE Conference on Evolutionary Computation 569-573
J A Joines and C R Houck 1994 On the use of non-stationary penalty functions to solve
nonlinear constrained optimization problems with GAs Proceedings of the First IEEE
Conference on Evolutionary Computation 579-584
K Juliff 1993 A multi-chromosome genetic algorithm for pallet loading Proceedings of the Fifth
International Conference on Genetic Algorithms 467-473
R G Le Riche C Knopf-Lenoir and R T Haftka 1995 A segregated genetic algorithm for
constrained structural optimization Proceedings of the Sixth International Conference on Genetic
Algorithms 558-565
G E Liepins M R Hilliard J Richardson and M Palmer 1990 Genetic algorithm applications to
set covering and traveling salesman problem in OR/AI: The Integration of Problem Solving
Strategies
12
G E Liepins and W D Potter 1991 A genetic algorithm approach to multiple-fault diagnosis in
Handbook of Genetic Algorithms (L Davis editor) Van Nostrand Reinhold New York
Z Michalewicz 1992 Genetic Algorithms + Data Structures = Evolution Programs Springer-
Verlag Berlin
Z Michalewicz 1995 Genetic algorithms numerical optimization and constraints Proceedings of
the Sixth International Conference on Genetic Algorithms 151-158
Z Michalewicz and C Z Janikow 1991 Handling constraints in genetic algorithms Proceedings of
the Fourth International Conference on Genetic Algorithms
A L Olsen 1994 Penalty functions and the knapsack problem Proceedings of the First IEEE
Conference on Evolutionary Computation 554-558
D Orvosh and L Davis 1994 Using a genetic algorithm to optimize problems with feasibility
constraints Proceedings of the First IEEE Conference on Evolutionary Computation 548-553
A L Patton W F Punch III and E D Goodman 1995 A standard GA approach to native protein
conformation prediction Proceedings of the Sixth International Conference on Genetic
Algorithms 574-581
D Powell and M M Skolnick 1993 Using genetic algorithms in engineering design optimization
with non-linear constraints Proceedings of the Fifth International Conference on Genetic
Algorithms 424-430
C R Reeves 1993 Modern Heuristic Techniques for Combinatorial Problems John Wiley & Sons
New York
J T Richardson M R Palmer G Liepins and M Hilliard 1989 Some guidelines for genetic
algorithms with penalty functions Proceedings of the Third International Conference on Genetic
Algorithms 191-197
M Schoenauer and S Xanthakis 1993 Constrained GA optimization Proceedings of the Fifth
International Conference on Genetic Algorithms 573-580
H-P Schwefel 1995 Evolution and Optimum Seeking John Wiley & Sons
W Siedlecki and J Sklansky 1989 Constrained genetic optimization via dynamic reward-penalty
balancing and its use in pattern recognition Proceedings of the Third International Conference on
Genetic Algorithms 141-150
A E Smith and D M Tate 1993 Genetic optimization using a penalty function Proceedings of the
Fifth International Conference on Genetic Algorithms 499-505
D M Tate and A E Smith 1995 A genetic approach to the quadratic assignment problem
Computers and Operations Research 22 73-83
D M Tate and A E Smith 1995 Unequal area facility layout using genetic search IIE
Transactions 27 465-472
S R Thangiah 1995 An adaptive clustering method using a geometric shape for vehicle routing
problems with time windows Proceedings of the Sixth International Conference on Genetic
Algorithms 536-543