Steven Gustafson | |||||||||||||||||||||
|
ResearchMy research has focused on analysing the genetic programming metaheuristic method. Genetic programming is an artificial intelligence paradigm used for automatic programming, model induction and scientific discovery. I also carry-out research in robotics (studying scaling and sensor issues), bioinformatics (protein folding using memetic algorithms), multi-agent systems (learning strategies for the keep-away soccer problem and the theory of memetics/cultural evolution) and analytical methods for understanding complex data (visualisation and complexity measures).
| |||||||||||||||||||||
Investigating measures of population diversity, how diversity correlates with performance and whether ideal levels of diversity measures exits.
See
Recent research examines issues with scaling and sensors in multi-robot
systems, to be presented at RoboSphere 2004.
See 
Researching how population diversity effects other aspects of search. In particular, how population diversity is effected by problem difficulty and the effects on solution size. A study using Lineage Selection shows how increasing genetic diversity can lead to smaller solutions. Also looking at how genetic programming samples the solutions space.
See
Researching the possibility of the N-Prisoners Puzzle (i.e. extension of the 3-Hats Problem) as a testbed domain. I am also interested in the possibility of logic mazes and related puzzles for machine learning testbeds, including the Keep-Away Soccer problem. Dissertation research proposed the Tree-String problem, a constructed problem combining structure and content objectives that should be useful for representing real-world problems.
See 
Extending a recent technique for tree visualisation to visualisation of a population in 3-dimensions. This research lead to new and novel measures for populations. Currently looking at other ways to visualise program dynamics when realized on tree structures.
See 
Currently researching various measures of population complexity, e.g. distance distributions and multidimensional scaling, to improve the understanding of population-based search methods.
Initially examined the behaviour represented by different solutions using
behaviour measures incorporating an aspect of complexity. See Report (coming soon),
Investigating the hybridisation of evolutionary algorithms with self-adaptive local search in the framework of Dawkins' ``memes''. This is part of an ongoing collaboration led by Natalio Krasnogor that looks at memes which are themselves ``evolvable'' and self-adaptive.
See
Developing intermediate representations of population information to allow efficient measures and visualisations for analysis of complex spaces. This is part of an ongoing collaboration led by Aniko Ekart.
See 
Proposed, simultaneously with Peter Stone, this toy problem as a machine learning testbed. Initial investigations used genetic programming in a hierarchical learning framework to develop strategies to play keep away soccer. Intend to extend this research at a later date.
See