The ISYS Project

Applying the Immune System Metaphor to Data Analysis and Machine Learning

This page is now out of date: contact smg@aber.ac.uk for further information.

Introduction

The ISYS project is an EPSRC funded project, in collaboration with Integral Solutions Ltd.. The research concerns a number of issues, shown below and is inspired by the human immune system and more specifically by the Immune Network Theory as a means to achieve immunological memory.
Issues :
  • The development of the understanding of artificial immune systems (AIS)
  • Applying AIS to real problems. So far the research has considered machine learning and visualisation. Two systems have been developed over the past 3 years for learning, the first was a system that would recognise promoter sequences in DNA strands and the second (and more recent research), identifying patterns and trends in mortgage application data. Using the immune network for the visualisation of similarities between classes in Java has also been done.
  • Developing AIS theory. Although the last 3 years have lead to some very useful work, it is now felt that some time should be spent on defining what is meant by an Artificial Immune System, in the context of data analysis and data mining. The current thread of research has seen a re-designing of the AIS algorithm, both from a theoretical view point and implementation view point. We are currently testing on the well known iris data set, see below for an example figure and hope very soon to have a version available of the software for you to play around with and give feedback from. Please contact Jon Timmis for more details.
    •

    The following figure presents the network viewer developed as part of the ISYS project. This figure illustrates a network generated from a mortgage application data set. By visualising he immune network, it is possible to see very easily areas of similarity in the network. The network captures trends in the data, as opposed to being a direct representation of the data itself. The cases in the evolved network may well be some of the training data, but some are also mutated versions of the training data and an initial network that was used to build the network with.

    Network viewer gif

    Further information on the project

    ISYS Publications

    A list of publications is available here. These are not available on line (thanks to Word producing dodgy postscript). However, if you email Jon Timmis he will send you copies of our papers.

    Future Research

    Research is focusing in the following areas. Much of this work will be continued by Jon Timmis as part of a PhD, which is being undertaken to further artificial immune system theory and applications, as the ISYS project is due to finish in May 1999.
  • The defining of what exactly an Artificial Immune System is, in the context of data analysis, data mining and machine learning. The past three years have seen a system develop that works, but it is time to fully define what we have learnt, how it all really fits together and so on. This will take the form of an in-depth paper describing the full workings of the system making close analogies to the human immune system. A new implementation of the immune network algorithm will be produced, that will be used as a comparative study with other data mining and cluster analysis techniques, using a variety of data sets. This is a major piece of work, but essential to the future of the viability of AIS.
  • Statistical analysis of the performance of the ISYS software.
  • The exploration of more immunity based concepts, these include self-recognition, T cells, histo-compatibility operators and exploring mutation in more depth.
  • The development of hierarchical immune networks. The network lends itself ideally to be transformed by hierarchical cluster analysis. This will generate a hierarchical immune network that could be used as an aid to the learning side of the system and also greatly improve the visualisation of the evolved immune network. A working algorithm has been developed and experiments are being run at this time.
  • This work leads onto thinking about incorporating hierarchical immune networks into the learning, by generating these hierarchical structures on the fly. This will be on top of the basic underlying algorithm, as it is a deviation away from immune system metaphors.
  • More we are sure along the way.
    •

    Personnel involved in the project

    Dr. Mark Neal / Mr. Jon Timmis,

    Centre for Intelligent Systems,
    Computer Science Department,
    University of Wales, Aberystwyth,
    Penglais Campus,
    Aberystwyth, Dyfed. SY23 3DB.
    Trendy Mail Pic mjn@aber.ac.uk,jot@aber.ac.uk
    23rd Feb. 1999