Improvement Research On Feedforward Neural Network Based On An Improved Bat Algorithm

To solve complex problems without feature of gradient,continuity and convexity,Traditional optimization methods cannot meet the requirements of computing time and space.Swarm intelligence algorithm which shows self-organization,cooperation and highly parallel features provides ideas for the problems.Bat algorithm is a new heuristic algorithm derived from swarm intelligence.It can be used to solve optimization problems without mathematical information.Bat algorithm has a faster convergence rate for solving the object problem.Through the study of bat algorithm,we discovered the defects.The first,a large number of bats leap out of solution space during the process of optimization.But the bat algorithm does not have a good strategy to allocation favorable locations for the individuals.Second,bat algorithm has a phenomenon of population gathering seriously,and it is lack of population diversity.Once the algorithm is trapped in a local optimal solution,it is difficult to diverge and escape from the attraction of the local attractor for bats.Third,the loudness and pulse frequency of the sound wave emitted by bats are inconsistent when the bats approach the target.It reduces the bat's ability to find new and better solutions.To solve the problems existing in bat algorithm,we control the bats that cross the solution border by relocation strategy,and we pull back the bats into the solution space to assign new positions randomly for them.Then,we use the mutation strategy when the population was trapped into stagnated and precocious status,which can make the bats search around with the current optimal solution as the center.Besides,we update the loudness and pulse frequency of the sound waves emitted by bats through liner gradation strategy,so that the changes are compatible with the process of algorithm optimization.In order to verify the effectiveness of the improved algorithm,we compared the optimization capabilities of the new algorithms and other algorithms from different solution space locations.The experimental results show that the new algorithm shows a stronger global search ability in solving multimodal and high-dimensional problems,and it has stronger robust.Training feedforward neural network can be seen as an optimization problem.There are many defects in back propagation method which updates network weights according the gradient of errors.For instance,larger networks have slow convergence.And the network's ability to approximate the optimization problem relies heavily on the initial weight.Moreover,the generalization ability of the network is inadequate.Bat algorithm shows outstanding performance in solving complex,high-dimensional optimization problems.This paper combined the improved bat algorithm and error back propagation algorithm and proposed a new algorithm for training the feedforward neural network.We compared the approximation ability of the networks trained by the new algorithm and error back propagation algorithm to verify the effectiveness of the new method.Finally,the network model trained by the new algorithm is applied to the research of classification problems.The results show that the proposed network model has stronger generalization ability.