Dr Cara MacNish, cara@csse.uwa.edu.au

Cara has studied Electrical Engineering (UWA), Artificial Intelligence (University of Cambridge, UK) and Psychology (UWA) and has held lecturships in Computer Science Departments at the University of York (UK) and at UWA since 1992. She has been working in Artificial Intelligence - that is, engineering biologically-inspired solutions to computational problems - and allied disciplines for almost 20 years.

The aim of these projects is to provide you with interesting technical problems (which you will want to write about in your thesis) combined with the satisfaction of building something that you can see working (which you will want to show off in your presentation).

Projects

1. Optimise This!

   Imagine you are given a black box with n inputs and one output. There is a very complex (unknown) relationship between the inputs and the output. Your job - find the inputs that optimise (maximise or minimise) the output. How would you do it?

   A more familiar example is shown on the right. Here the inputs are x and y co-ordinates, and the aim is to find the deepest (or alternatively highest) point in the landscape. In this case it happens to be a random fractally generated landscape, so it is not possible (practically speaking - you would have to exploit the number limitations of the computer) to find the exact deepest point. The problem then becomes finding the best possible solution with a given computational resource.

   This problem arises in numerous computing (and non-computing) applications - in fact most areas of life. Imagine for example that the input variables represent the presence of geological features, and the output variable in the above landscape (rather than depth) is a probability measure of finding oil. Or that the inputs represent fuel/air ratios in a fuel injection system and gear ratios in a drive train, and the output is the fuel efficiency of a car. Or that the inputs represent the values of international stock market indices, and the output represents the expected rise or fall in the Australian index. There are infinitely many examples.

   In general, you cannot be sure you have the right answer without testing all (possibly infinitely many) combinations of inputs, and even that would only work if the black box was deterministic. Luckily, however, real world systems tend to have some kind of structure that can be exploited.

   Global optimisation techniques such as evolutionary algorithms and particle swarm optimisation attempt to adapt to and exploit the underlying structure of complex problems. While they have been very successful on some problems, there are many for which the standard algorithms fail to find good solutions in a realistic computational time. The aim of these projects is to improve the performance of optimsation algorithms by combining global techniques with local search (a combination sometimes referred to as memetic algorithms).

   These projects will make use of the Huygens Benchmarking Suite. You may also like to enter the CEC2006 Huygens Probe Competition.

   Illustrative Past Projects

   Tsou, D., The Particle Swarm Optimiser and Recurrent Neural Networks, Honours Thesis, CSSE, UWA, 2002

   Di Pietro, A., Adaptive Learning Techniques for Opponent Modelling, Honours Thesis, CSSE, UWA, 2000

2. Adaptive Approaches to Object Tracking in AIBO Robotic Dogs
   (Additional advice will be available in 1st semester from Dr Wei Liu.)

   Automatically tracking objects in visual images (as well as other kinds of sensory data such as sonar) is important in many applications. Examples include air and naval traffic control, security and defense applications, autonomous robotics (for example RoboCup, or following a human) and interpreting hand movements in sign languages (such as Auslan).

   There are many factors that make object tracking interesting - and difficult. These include shape deformation (imagine the changes in apparent shape that take place as a paper aircraft curves towards you, or the changes in hand shape as someone communicates in sign language), occlusion (when the object passes behind another), and many sources of noise (for example lighting conditions, sonar reflections, etc).

   The aim of these projects is to use adaptive techniques (eg. neural networks, evolutionary algorithms) to develop methods of object tracking that are robust to reasonable levels of deformation, noise, and limited occlusion, and to deploy the resulting algorithms on the Aibo robotic dog.

   Example References

   Kumazawa, I., "Target tracking by matching a shape represented by a tree of sigmoid functions", Pattern Recognition Letters, 21, 661-675, 2000

   

   Nistico, W. and Rofer, T., "Improving Percept Reliability in the Sony Four-Legged Robot League".

   Freeston, L., "Applications of the Kalman Filter Algorithm to Robot Localisation and World Modelling".

   Illustrative Past Projects

   Cristobal, T., Discrimination of a Hand from a Face using Template Matching, Honours Thesis, CSSE, UWA, 2003