abstract = "I explore several extensions to genetic programming
for applications involving the forecasting of real
world chaotic time series. We first used Genetic
Symbolic Regression (GSR),which is the standard genetic
programming technique applied to the forecasting
problem in the same way that it is often applied to
symbolic regression problems [ Koza 1992, 1994]. We
observed that the performance of GSR depends on the
characteristics of the time series, and in particular
that it worked better for deterministic time series
than it did for stochastic or volatile time series.
Taking a hint from this observation, an assumption was
made in this study that the dynamics of a time series
comprise a deterministic and a stochastic part. By
subtracting the model built by GSR for the
deterministic part from the original time series, the
stochastic part would be obtained as a residual time
series. This study noted the possibility that GSR could
be used recursively to model the residual time series
of rather stochastic dynamics, which may still comprise
another deterministic and stochastic part. An algorithm
called GRR (Genetic Recursive Regression) has been
developed to apply GSR recursively to the sequence of
residual time series of stochastic dynamics, giving
birth to a sequence of sub-models for deterministic
dynamics extractable at each recursive application. At
each recursive application and after some termination
conditions are met, the submodels become the basis
functions for a series-expansion type representation of
a model. The numerical coefficients of the model are
calculated by the least square method with respect to
the predetermined region of the time series data set.
When the region includes the latest data set, the model
reflects the most recent changes in the dynamics of a
time series, thus increasing the forecasting
performance. This chapter shows how GRR has been
successfully applied to many real world chaotic time
series. The results are compared with those from other
GSR-like methods and various soft-computing
technologies such as neural networks. The results show
that GRR saves much computational effort while
achieving enhanced forecasting performance for several
selected problems.",