Design optimisation of the Nadaraya-Watson fuser using a genetic algorithm

The Nadaraya-Watson (N-W) statistical estimator based on Haar kernels can be used to implement a fuser based on empirical data. Fuser design essentially consists of the following interrelated activities: select a set of n observations from a pool of p prior observations; select a value for the bandwidth. Optimal fuser design can therefore involve a very large search space. This paper proposes the use of a genetic algorithm (GA) to optimise the fuser design. The GA is used to evolve optimal values for the bandwidth and subset of observations used to implement the fuser. Indicative test results are provided. The N-W fuser is shown to perform better than the best single sensor. The GA provides better results than manual design optimisation, with the performance of the N-W fuser comparable to that achieved using a feedforward neural network.