The Research On The Neural Network-Based Mobile Robot Control

Behavior-based mobile robot control tightly couples sensory inputs and effecter outputs, to allow the robot to quickly respond to changing and unstructured environments by a simple structure and has a good performance in the character of real time.However, the learning process of behavior-based mobile robot is a kind of designing learning, not a self-learning. Its limitations include the robot's inability to have memory, internal representations of the world, or the ability to learn over time. The robots are unable to change their actions with varying of environments.It is crucial that a robot should have both learning and evolutionary ability to adapt to dynamic environments. This thesis proposes a new robot behavior decision controller using modified Elman Neural Network (Elman NN) .The Elman NN has the advantageous time series prediction capability because of its memory nodes, as well as local recurrent connections. The modified Elman NN controller improves not only the sensibility of historical data, but also the ability of adaptation in the dynamic environments.The training algorithm for the Elman neural network is similar to the BP learning algorithm, as both based on the gradient descent principle. However, local minima caused by the regular BP learning algorithm often results in an unavoidably large approximation error that may reduce NN prediction accuracy. Therefore the Genetic Algorithm (GA) is introduced in this thesis in order to optimization of the weights in the Elman neural network controller to achieve better behavior performance. This king of training algorithm is called Evolved BP algorithm.The behavior-based mobile robot always faces unknown tasks and environments. Therefore the robot should have the ability of self-learning, which means the robot is able to modify its actions according to the changing of tasks and environments. This thesis presents the SARSA-Reinforcement Learning (unsupervised learning) in the experiment of obstacle avoidance and wandering. The computer simulation is given to show the validity of the method.