Algorithm animation and its application to artificial neural network learning

Algorithm animation is a means of exploring the dynamic behavior of algorithms using computer-generated graphics to represent algorithm data and operations. Research in this field has focused on the architecture of flexible environments for exploring small, complex algorithms for data structure manipulation. This thesis describes a project examining two relatively unexplored aspects of algorithm animation: issues of view design effectiveness and its application to a different type of algorithm, namely back-propagation artificial neural network learning. The work entailed developing a framework for profiling views according to attributes such as symmetry, regularity, complexity, etc. This framework was based on current research in graphical data analysis and perception and served as a means of informally evaluating the effectiveness of certain design attributes. Three animated views were developed within the framework, together with a prototype algorithm animation system to run each view and provide the user/viewer interactive control of both the learning process and the animation. Three simple artificial neural network classifiers were studied through nine structured investigations. These investigations explored various issues raised at the project outset. Findings from these investigations indicate that animated views can portray algorithm behaviors such as convergence, feature extraction, and oscillatory behavior at the onset of learning. The prototype algorithm animation system design satisfied the initial requirements of extensibility and end-user run-time control. The degree to which a view is informative was found to depend on the combined view design and the algorithm variables portrayed. Strengths and weaknesses of the view design framework were identified. Suggested improvements to the design framework, view designs and algorithm system architecture are described in the context of future work