Genetic Evolution of Symbolic Signal Models
Proceedings of the Second International Conference on
Natural Algorithms in Signal Processing, NASP'93, Essex University, UK, (15-16 November 1993)Zusammenfassung
This paper reports on a novel method of signal
modelling that employs a variable model structure as
opposed to the fixed model structure used in
conventional methods. The functional form of the model
along with any required numerical parameters are
simultaneously estimated from the signal sequence to be
modelled. This is accomplished by defining the model
functional and its parameters in terms of structured
lists of symbols, and using an estimation algorithm
that can infer symbol lists from the given data.
Motivated by the recent work in cellular coding and
evolutionary computation, we use Genetic Programming
(GP) to evolve high quality model structures. This is
based on a coding of the model in terms of an
expression tree in polish form which can then be
manipulated and optimised using standard Genetic
algorithm (GA) techniques. In conjunction with the
model structure evolution, we use Simulated Annealing
to optimise the numerical parameters of the model and a
set of production rules to minimise the model order.
The paper discusses how these three processes can be
combined to yield a powerful general purpose modelling
system