Time series forecasting based on aggregation of interval type-2 fuzzy
logic systems
O.V. Cheboraka
A problem of time series forecasting is encountered in many fields of
science, engineering, economy, medicine, industry, agriculture. Traditional
models usually generate only a single forecast value. But uncertainties being
present in subject field prevent from obtaining sufficiently accurate single
forecast value. A powerful tool of uncertainty handling is type-2 fuzzy logic. General
type-2 fuzzy logic systems are computationally intensive due to the complexity
of type reduction. Therefore I design interval type-2 fuzzy logic systems for forecasting
time series due to their good capabilities of uncertainty handling and less computationally
intensive process. In this paper I study an aggregating interval type-2 fuzzy
model for forecasting time series. This model consists of a certain number of
interval type-2 fuzzy logic systems and an aggregator used for calculating a resultant
prediction interval. The interval type-2 fuzzy logic systems differ in their
inputs number. Each interval type-2 fuzzy logic system is obtained from a type-1
fuzzy logic system trained previously by a genetic algorithm. I use an
iterative procedure to get interval type-2 fuzzy logic system parameters from the
type-1 fuzzy logic system. I also use the genetic algorithm for further training
of every interval type-2 fuzzy logic system. To select which interval type-2
fuzzy logic systems should be included to the aggregating
interval type-2 fuzzy model I apply root mean square width of prediction
interval on testing samples as a selection criterion. I introduce a formula for
aggregation of prediction intervals. Finally, I design the aggregating interval
type-2 fuzzy model to predict number of patients coming to a health resort. The
simulation demonstrates better generalization capabilities of the aggregating
interval type-2 fuzzy model that outperform those of single interval type-2
fuzzy logic systems.