AI Group Research Overview
|
Language and Logic
|
|
One of the many achievements of coordinate geometry has been to provide a conceptually elegant and unifying account of spatial
entities. According to this account, the primitive constituents of space are points, and all other spatial entities---lines curves,
surfaces and bodies---are nothing other than the sets of those points which lie on them. The success of this reduction is so great
that the identification of all spatial objects with sets of points has come to seem almost axiomatic. For most of the previous
century, however, a small but tenacious band of authors has suggested that more parsimonious and conceptually satisfying
representations of space are obtained if we adopt an ontology in which regions, not points, are the primitive spatial entities.
These, and other, considerations have prompted the development of formal languages whose variables range over certain subsets (not
points) of specified classes of geometrical structures. We call the study of such languages mereogeometry. In the past decade, the
Computer Science community in particular has produced a steady flow of new technical results in mereogeometry, especially concerning
the computational complexity of region-based topological formalisms with limited expressive power.
|
|
|
Language engineering: building software tools to support people doing useful things with natural language text.
The MultiFlora project aims to provide proof of concept that document information extraction can be improved by the
analysis of multiple paralell texts. Applied to botanical taxon descriptions, we believe this technique has the potential
to be a useful tool in biodiversity informatics. Multiflora aims to (i) ground the work in a structured domain model,
capitalising on the experise in Medical Informatics at Manchester, and (ii) extend taxonomic coverage, contributing significantly
to biodiversity research.
|
|
Machine Learning
|
Multiobjective optimisation in unsupervised and semi-supervised learning
Contact: Joshua Knowles
This research, carried out with PhD student, Julia Handl, investigates the use
of explicit multiobjective optimization methods in exploratory data analysis.
We recognised that data clustering and feature selection performance is often
defined in terms of external measures (not available to the learner), but
that only a single internal criterion is used by the learner as a proxy. If
this criterion is not perfectly "aligned" with the external criteria, poor
performance results. By, instead, optimising over several internal criteria
simultaneously, we found the learner is able to recognise the structure
inherent to the data more easily (from the shape of the tradeoff front) and
give more accurate solutions. We have observed that our multiobjective
clustering algorithms, which include a model selection step (i.e. automatic
selection of the most appropriate solution) perform better than
state-of-the-art ensemble clustering approaches, on many problems, and we've
explained why this is the case. The multiobjective approach can also smoothly
incorporate extra knowledge, as in the semi-supervised learning case, where a
small amount of external class information is available.
|
|
|
We use probabilistic methods to propagate the uncertainty of gene expression levels from probe-level analysis of Affymetrix
microarray data to higher level of gene expression data analysis. The focus of the project is twofold: developing tractable
probabilistic models for microarray data, allowing reliable estimates of the experimental noise associated with each gene
and experiment. At a higher level, we aim to produce algorithms and software for data analysis (such as PCA, clustering,
etc..) that uses the information obtained in the probe-level analysis. This would allow to obtain uncertainty levels on the
final outcome of the analysis, and hopefully it should give a principled way to automate many of the heuristic procedures
currently used in microarray data analysis.
|
|
|
PHASE is a package that performs molecular phylogenetic inference. The software seeks to accurately compare molecular
sequences to determine the likely evolutionary relationships between a group of species. It is designed specifically
for use with RNA sequences that have a conserved secondary structure, e.g., rRNA and tRNA.
|
|
|
Agent-Based Computational Economics (ACE) is a rapidly expanding approach for modelling financial markets, holding
both explanatory power and strong modelling capabilities.
Time representation in ACE models is an issue, since financial markets like most human activities, evolve in continuous
time; however many modelling approaches discretize this, to simplify calculations.
We are investigating the properties of a discrete-event versus continuous-time Artificial Stock Market. The aims are (i) to determine the
conditions in which traditional ACE models remain valid (or break down) once transferred to continuous time, and (ii) to
propose a model highlighting the effects of synchronisation and information dissemination in financial markets.
|
|
Ensemble Learning
Contact:
Gavin Brown
In collaboration with researchers at the University of Birmingham, we have analysed and
enhanced ensemble learning methods for regression and classification problems.
Our primary focus has been on understanding the issue of diversity in ensemble systems.
We have found by taking
the combination mechanism for the ensemble into account we can derive an error
function for each individual learning machine that balances ensemble diversity with individual
accuracy. We have demonstrated that these methods control the bias-variance-covariance trade-off
systematically, and can be utilised with any estimator that can be cast in a generalised linear
regression framework, for example MLPs and Radial Basis Function networks.
|
|
Estimation of Distribution Algorithms
Contact:
Jon Shapiro,
Estimation of distribution algorithms (EDAs) use machine learning techniques to solve optimization problems, by trying to learn the
locations of the more promising regions of the search space. In particular, a probabilistic model, such as a graphical model, is used
to generate candidate solutions, and learning is used to adapt the probability model to explore more promising regions of space.
Most of the research in this field has been focused on producing new algorithms based around new probability models or learning
rules. Our research is focused on understanding how they work, why they can fail, and based on the previous two points, how they can
be improved.
|
|
In collaboration with Dr. Adam Prügel-Bennett of Southhampton University, we
have developed a formalism for describing genetic algorithm evolution. This is based on the dynamics of cumulants of a phenotypic
trait in a population, and uses maximum entropy inference and statistical mechanics techniques. This has been shown to predict the
dynamics of a finite-sized population very accurately on a number of problems.
The important question are: Can this method be used to predict effective learning parameters, such as the mutation and crossover
rates? Can this method be used to test the effectiveness of different representations and approaches?
|
|
Robotics
|
Advances in the state of the art of Urban Search and Rescue (USAR) robotics have come about from the associated
progression in the general robotics field and the need for autonomous/semi autonomous systems that can rapidly explore
confined and structurally unstable spaces in collapsed buildings to locate trapped human beings. Current deployable USAR
robots have achieved only limited success and have suffered a range of problems such as traction problems, communications
failure, sensor failure, control system problems or even simply becoming trapped themselves.
Many current USAR robots treat debris as an obstacle to be avoided, i.e. navigated around. The chaos and cramped
conditions present in most collapse situations cannot allow this to be assumed and robots are necessary that interact with
the debris field. Although some robots are able to some extent drive over debris, this is not enough and robots must be
equipped with the ability to interact with and manipulate the debris to progress further into a collapse site with the aim
of locating survivors.
Building on the current state of the art of USAR research, the group is looking at producing an advanced intelligent robot
with dextrous manipulators. To this end, the group believes that important basic knowledge such as the characteristics of
debris must first be understood.
This work is supported by EPSRC GRANT: EP/C510097/1
|
|
Intelligent Robotics
Contact: Rob Richardson
If robots are to behave autonomously then they must have the capability of independent movement and 'intelligent' control
strategies. Manchester robotics has a long history of developing adaptive 'intelligent' controllers for mobile robotic
systems. These mobile robotic systems can be either locomoted by wheels, legs or alternative novel methods. Applications
of such robots vary from urban search and rescue to planetary exploration.
Robots that are in everyday contact with humans must be designed and controlled in a suitable manner. Interaction control
techniques are being developed to ensure robots are of no danger to humans. Moreover, on a more subtle level, these robots
must reassure and not intimidate humans. Manchester robotics is investigating robots that exhibit body language.
|
|
