Adaptive visual representations for autonomous mobile robots using competitive learning algorithms
McNeill, Dean K.
MetadataShow full item record
This thesis examines issues surrounding the class of unsupervised artificial neural network learning algorithms known as competitive learning. Four variations of competitive learning algorithms are presented and compared, both theoretically and based on their relative performance in the solution of a number of low and high dimensional input environments. In particular, the thesis discusses efficacy of these algorithms in learning appropriate representations of visual information in robots Comparisons of hard competitive learning (HCL) and soft competitive learning (SCL) in the low dimensional discrimination of Gaussian data clusters showed that SCL consistently produces superior solutions. As well, the tendency of HCL to become trapped in sub-optimal solutions was analysed and found to be an inherent shortcoming of the winner-take-all nature of the algorithm. It was also found that selection of an appropriate network size may be achieved through the use of a simple pruning technique if a surplus of network units are provided to begin training. Further investigations involving HCL, SCL, and both the DeSieno and Krishnamurthy implementations of frequency sensitive competitive learning (FSCL) show that the latter (FSCLK) produces the most consistently reliable solutions to a number of learning tasks. This result was obtained as a consequence of extensive testing involving a high dimensional data clustering problem. That problem concerned the adaptive identification and classification of motion via an array of optical sensors residing on an autonomous mobile robot. The selection and arrangement of sensors used by this robot were derived from the vision system of jumping spiders. Operation of an integer-only version of FSCLK on the actual robotic hardware demonstrates the system's ability to cluster some aspects of the motion identification task. The inability to completely identify and generalize to novel input patterns is attributed to deficiencies in the sensors used and is not an inherent shortcoming of the algorithm. These deficiencies can be corrected through the use of some preprocessing of the raw sensor readings. As well, during the course of this study the winner-take-all activations used by the frequency sensitive algorithms were replaced with analog activations, resulting in significantly improved network generalization.