Competitive Learning in Robust Communication

A modern communication system should be accurate, reliable, robust and make efficient use of the available channel. Vector quantization is beginning to prove its effectiveness for source coding in applications such as cellular mobile telephony. The design of vector quantizers is typically done by clustering methods. A vital part of the design is to incorporate robustness against channel disturbances which may be accomplished by carefully choosing the codewords transmitted on the channel. A joint optimization of the source and channel codes, can be made superior to a separate design of the coders. <p> This thesis describes non-redundant source/channel coding. Joint source/channel optimization as well as separate optimization of the index assignment are investigated. Algorithms for joint source/channel optimization are developed using a sample-recursive approach. <p> Although the index-assignment problem is NP-hard, several procedures are developed to find close to optimal solutions. Hypercube mapping is introduced as a convenient way to describe vector quantizers. A new method for index assignment has been developed as a direct consequence of this description. A conceptually attractive, compact, exact analytic expression is derived for the end-to-end distortion for maximum entropy encoders using transmission on binary symmetric channels. Theoretical bounds for the end-to-end distortion are derived. Conditions for reaching the bounds are presented in terms of the linearity of the hypercube mapping. <p> The theoretical analysis is verified by extensive simulations using Gaussian sources. A database of 250 vector quantizers is made available for public use. <p> The thesis also concerns iterative algorithms for solving classification problems. Model-based methods as well as methods based on artificial neural networks are investigated. A new fast method has been developed for training a model-based classifier. A new adaption rule is also proposed for Kohonen Feature Maps, which converges faster and improves performance, including recognition rate, compared to previous methods