Publicação
Artificial intelligence in knowledge management for Time Series Forecasting
| Resumo: | Knowledge Management (KM) is a keen topic for an organization, in particular to those that have to deal with knowledge acquired from different sources, either from its own experiences or from that of others, to decide on the effective use of that knowledge to fulfill the goals of the organization. As representative examples of KM, one may have the object-oriented data bases, hypermedia or concept maps. On the other hand, techniques developed in Artificial Intelligence for knowledge representation and discovery may be of great use in KM; in particular, it seems natural to explore the potential of the organization past data to deal with management decisions of the present. One way is to use Time Series Forecasting (TSF), where forecasts are based on pattern recognition of past observations ordered in time. Traditional TSF methods, such as the Holt-Winters and the Box-Jenkins ones, are based on particular characteristics of the Time Series (TS), such as trend or seasonal effects. These methods work with well behaved TS, but present some drawbacks on TS with noise or some unknown nonlinear relations among the TS data. An alternative approach is the use of Artificial Neural Networks (ANNs), which present two main advantages: ANNs can extrapolate patterns from past data, even in TS with noise, and may adapt their behavior as new data comes in. A problem with the use of this approach is the search time for the best ANN architecture, which involves a large searching space, demanding a huge computational effort. Other aspect of concern is that of TS data filtering. Not all lags of the TS have the same influence over the forecast. Feeding the ANN with a big time window will slow the ANN forecasting efficiency. To solve these pitfalls, one may use random search, hill climbing or genetic procedures. The last ones are known to work well on problems of combinatorial nature, obtaining good solutions where other methods seem to fail. This paper presents an integrated approach for TSF: a set of rules will create the training cases, based on some lags of the TS; these rules and the ANN parameters will be encoded on the genetic chromosomes; finally, each ANN will be trained, leading to competition. |
|---|---|
| Autores principais: | Neves, José |
| Outros Autores: | Cortez, Paulo |
| Assunto: | Neural networks Time series |
| Ano: | 1997 |
| País: | Portugal |
| Tipo de documento: | comunicação em conferência |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | Knowledge Management (KM) is a keen topic for an organization, in particular to those that have to deal with knowledge acquired from different sources, either from its own experiences or from that of others, to decide on the effective use of that knowledge to fulfill the goals of the organization. As representative examples of KM, one may have the object-oriented data bases, hypermedia or concept maps. On the other hand, techniques developed in Artificial Intelligence for knowledge representation and discovery may be of great use in KM; in particular, it seems natural to explore the potential of the organization past data to deal with management decisions of the present. One way is to use Time Series Forecasting (TSF), where forecasts are based on pattern recognition of past observations ordered in time. Traditional TSF methods, such as the Holt-Winters and the Box-Jenkins ones, are based on particular characteristics of the Time Series (TS), such as trend or seasonal effects. These methods work with well behaved TS, but present some drawbacks on TS with noise or some unknown nonlinear relations among the TS data. An alternative approach is the use of Artificial Neural Networks (ANNs), which present two main advantages: ANNs can extrapolate patterns from past data, even in TS with noise, and may adapt their behavior as new data comes in. A problem with the use of this approach is the search time for the best ANN architecture, which involves a large searching space, demanding a huge computational effort. Other aspect of concern is that of TS data filtering. Not all lags of the TS have the same influence over the forecast. Feeding the ANN with a big time window will slow the ANN forecasting efficiency. To solve these pitfalls, one may use random search, hill climbing or genetic procedures. The last ones are known to work well on problems of combinatorial nature, obtaining good solutions where other methods seem to fail. This paper presents an integrated approach for TSF: a set of rules will create the training cases, based on some lags of the TS; these rules and the ANN parameters will be encoded on the genetic chromosomes; finally, each ANN will be trained, leading to competition. |
|---|