Competitive Learning in Autonomous

A BSTRACT Artificial Neural Network (ANN) algorithms provide powerful techniques for the construction of computing systems which are capable of adapting to non-stationary or incompletely specified operating environments. Unsupervised learning is of particular interest in such situations. This paper reports the results of empirical studies into the characteristics of standard and soft competitive learning algorithms and the embedding of the algorithm in a simple microcontroller based system for the operation of an autonomous mobile robot. Standard plastic LEGO parts are used for the structural components of these robots with the neural algorithms providing the control of the system from an on-board MC68HC11 microcontroller. The inputs to the network are provided by various analog sensors mounted on the robot chassis, while the outputs control the operation of the drive motors. Results of tests conducted on the algorithms demonstrate that soft competitive learning provides more robust performance than winner-take-all learning and that this algorithm can be realized in a simple microcontroller system if approximations are made to the theoretical computations. These approximations involve replacing the exponential function with a look-up table of modest size. Robotic systems, which are most commonly used in large industrial applications such as automobile assembly lines, are both efficient and reliable tools for industrial manufacturing. These robots are designed and programmed to perform simple predictable repetitive tasks. However, any significant variation in the robot's environment, such as the change of vehicle model being manufactured, will require further intervention from a human operator in order to reprogram the robot's duties. This type of occasional intervention and setup is acceptable in industrial applications where thousands of identical items are being produced. Yet, there exist many other applications where the use of robots is highly desirable but the details of the operating environment make it difficult or impossible for a human operator to intervene in the robot's operation. One examples of such an environment is the exploration of planet surfaces by space probes. Even exploration of one of the closer planets, Mars, would require a worst case round trip command-response communication time of approximately 44 minutes. This would make control of a robotic probe extremely tedious. Clearly, an autonomously vehicle would provide a more efficient mechanism for planetary exploration. While space exploration is one of the more exotic applications, a number of equally challenging environments exist in our everyday lives. A number of groups, including primarily The Robotics Institute at Carnegie-Mellon …