Memetic Approaches for Optimizing Hidden Markov Models: A Case Study in Time Series Prediction

We propose a methodology for employing memetics (local search) within the framework of evolutionary algorithms to optimize parameters of hidden markov models. With this proposal, the rate and frequency of using local search are automatically changed over time either at a population or individual level. At the population level, we allow the rate of using local search to decay over time to zero (at the final generation). At the individual level, each individual is equipped with information of when it will do local search and for how long. This information evolves over time alongside the main elements of the chromosome representing the individual. We investigated the performance of different local search schemes with different rates and frequencies as well as the two newly proposed strategies. Four time series of the exchange rate were used to test the performance. The results showed the inconsistent behaviour of the approaches that used manual settings on local search's parameters, while showing the good performance of adaptive and self-adaptive strategies.