The dynamical systems model for the simple genetic algorithm due to N'ose [12] can be simplified to the ca,se of zero crossover, and to fitness functions that divide the search space into relatively few equivalence classes. This produces a low-dimensional system for which the fixed-point can be calculated; it is the leading eigenvector of the system. This technique, applied elsewhere [11] to Royal Road functions, is adapted in this paper to apply to ftmctions of imitation. Infinite population fixed-points are calculated for some simple examples, including trap functions that have previously been analysed in terms of deception [2]. The effects of eigenvectors outside the population space are explored, and finite population behaviour is examined in this way. Two surprising theorems are demonstrated for infinite populations; that every population distribution is the fi.xed-point for some fitness function; and that fitness functions exist for which evolution goes "backwards" towards the global minimum. Theoretical results are backed up with experiments using real, finite population genetic algorithms.
[1]
Michael D. Vose,et al.
Random Heuristic Search
,
1999,
Theor. Comput. Sci..
[2]
Melanie Mitchell,et al.
Finite populations induce metastability in evolutionary search
,
1997
.
[3]
James P. Crutchfield,et al.
Statistical Dynamics of the Royal Road Genetic Algorithm
,
1999,
Theor. Comput. Sci..
[4]
Adrian Thompson.
Evolving fault tolerant systems
,
1995
.
[5]
Kalyanmoy Deb,et al.
Analyzing Deception in Trap Functions
,
1992,
FOGA.
[6]
John J. Grefenstette,et al.
Deception Considered Harmful
,
1992,
FOGA.
[7]
Melanie Mitchell,et al.
Relative Building-Block Fitness and the Building Block Hypothesis
,
1992,
FOGA.
[8]
Siddhartha Bhattacharyya,et al.
General Cardinality Genetic Algorithms
,
1997,
Evolutionary Computation.
[9]
F. W. Kellaway,et al.
Advanced Engineering Mathematics
,
1969,
The Mathematical Gazette.
[10]
M. Eigen,et al.
Molecular quasi-species.
,
1988
.