Research On Recurrent Neural Network Based Approaches To Path Planning Of Mobile Robot

Based on the existing research outcome on path planning of mobile robot, this paper systematically discusses three aspects including structure, function and training algorithms of recurrent neural network, the combination of evolutionary algorithms with recurrent neural network, and the application of recurrent neural network in path planning of mobile robot.Considering the weakness about the complication, inefficiency and restriction of the traditional training approaches, the application of evolutionary algorithms to neural network design optimization is introduced. The fitness of a population is adjusted with simulated annealing algorithm in evolving process. The diversity of similar individuals will be enlarged during the posterior evolving process. So the superiority of excellent individuals is outstanding. The combination of Gaussian and Cauchy mutations is used to keep larger mutation step and escape from local minima. Crossover and mutation probabilities are adjusted automatically according to the varying diversity of population and the fitness of individuals. And different values are assigned to them reasonably during the evolving process to prevent premature and speed evolving convergence.Traditional artificial potential field method has unexpected local attracted points and can't guarantee the integrity of the path planning. A method of path planning is proposed based on recurrent neural network with self-feedback in this paper. The arrangement of the neurons coincides with the discretized representation of configuration space. The output of the neurons is relevant to the local potential field values. The target neuron has the maximal positive neural activity via appropriate external stimulation. The activities of the neurons, in obstacle fields and the local neighborhoods, are made to zero. A potential field with single apex is shaped in the whole configuration space. And robot can reach the destination with shortest path according to steepest ascent direction in dynamic and crowded environments..Traditionally, the methods to path planning of mobile robot have no adaptability and real-time ability. The robot can't learn from unknown environment. In this paper recurrent neural network is used as compute model of mobile robot controller. The mapping relation is constructed between input of sensors and output of actuator based on recurrent neural network with its high learning and dynamic nonlinear modeling ability. The recurrent neural controller is trained via improved evolutionary algorithm. The robot can obtain adaptive behavior automatically from learning process. Two different neural controllers are evolved in similar environment. Their dynamic performance is compared each other. In this way we try to choose more appropriate neural controller structure.