Implementing Decision-Making Methods Based on Multipe Neural Networks

This paper presents an approach to implementing a learning and decision-making method for mobile robots in manufacturing environments based on multiple neural networks. Usually, these networks are used for strongly separated architectures, but in our research we used them to diversify the similarities in reading environments caused by low resolution sonars, low accuracy in motion of drive train and actuators. These networks work independently in their own domain, while a decision network brings the result. We have tested different configurations of neural networks, but best results were obtained with multiple neural networks. As an experiment to support our research we have considered a technology environment that could be improved by use of mobile robots. Given the high costs of building a real full-scale mobile robot, we have decided to down-size the problem and evaluate this possibility with LEGO Mindstorms NXT and neural networks using Matlab. The motion of the robot, determined by a choice the robot has to make, was modeled and implemented by a software solution, set up with accuracy that enables mapping, obstacle recognition and avoidance. A more qualitative evaluation can be obtained by taking hardware into consideration (quality of drive engines and actuators of the mobile robot, accuracy of signal processing from the ultrasound sensor, motion tolerance of the sensor head, etc.).