Discovering time-series building blocks using an artificial intelligence framework

This thesis contributes to the area of time-series prediction by presenting a novel, noise resistant, Artificial Intelligent time-series prediction framework called the "Significance Engine", created by combining together the techniques of Takens' Theorem and dynamic self-organising maps. The central concept that the system is built around, is that time-series can be broken down into both random and non-random components and that the non-random components can be used to help predict future time-series movement. It is a highly parallel system which combines the use of multiple Dynamic Self-Organising Map units of the 'Grow When Required' type with time delay embedding techniques. The framework works by maintaining a finite memory in which it builds a representation of the input waveform's distribution to a predetermined level of detail, after which it is able to selectively forget older less useful information making room for the new.' It learns incrementally, so the more data the system is fed from the input distribution, the better its internal model and the better the predictions it is thus able to make. After it has been initialised, and primed using historical data, it is able to recognise re-occurring patterns in highly noisy waveforms and to successfully make future predictions based on what occurred historically when these motifs (reoccurring patterns) were previously observed. Testing of the system was carried out initially using the well-known random-walk time-series, interspersed with varying degrees of a well-known motif, to show clearly how it performs under controlled conditions. Next, the system's predictive capabilities were benehmarked using the Mackey-Glass chaotic time-series, a reproducible waveform that is used in time-series research as a comparative base line.