Multilayer neural network with back propagation: hardware solution to learning XOR

摘要

In this paper we present our current work on a hardware implementation of Artificial Neural Networks (ANNs). In specific, we are working to create a multilayer neural network with back propagation that can learn the XOR training input. The larger agenda of this project is to make progress in intelligent robot controllers. This is a difficult issue because microcontrollers used as onboard controllers for the robot must be capable of dealing with a changing environment. ANNs are particularly suitable candidates for a control system; they can deal effectively with the micro-structure of the robot's sensors and motors, they are robust in affecting behavior, can adjust to an environment, have tolerance to noise, and are well structured for producing the appropriate behavior [1]. To compute the best neural network, genetic algorithms (GAs) are often used. Since they can be computationally expensive, GAs are usually carried out in a simulation, where the best neural network control program is evolved. Parker and Lee (2002) used a genetic algorithm in a simulation to learn the best neural network for an individual leg controller on a hexapod robot [2]. Floreano and Mondada (1994) allowed a neural-network driven robot to completely evolve on its own by directing its genetic evolution to a workstation through a wire while the robot was moving [1]. Despite this autonomy, the robot was still not completely autonomous because it relied on the workstation to perform the genetic algorithm. In addition, the ANN on the microcontroller was executed through a sequence of serial instructions, unlike the parallel method of a biological neural network.