Broadly my project interests can be divided in the areas of artificial intelligence, logic-related topics, and economics. Furthermore I have a particular interest in software tools, e.g., tools for manipulating/administrating digital pictures. Research wise I am interested in reasoning , heuristic search, and economic models.
MSc students should note that over large part of the summer I will not be in Birmingham. While supervision via Skype is possible, a particular degree of independent work is needed.
Software
AI related projects
2. Modelling of a Simple Eco-System
3. From Planning to Automated Programming
5. An Imitation System for Composing Music
Reasoning and Logic Related Projects
7. A Model Generator for First-Order Logic
9. Modelling Fallacies of Human Reasoning
Economics
Course Suitability: Any
Brief Description: Sudoku games on a computer typically give just
a computer version of the paper game, the same rules apply, no
particular advantage of the possibilities of a computer are
taken. It is proposed here to change the rules in a form that
additional rules and constraints are added, which are difficult
to match on paper, e.g. that the structure becomes unstable or
fades with increasing time.
Languages and/or other Software: Java
Prerequisite:
None
Course Suitability: Any
Brief Description: Many systems show stability only when the rules of
the game are well-balanced. For instance, a carnivore species may be
so successful that they eat up all their prey and then starve. In this
project a simulated environment (may be ecological or economical)
should be built and tests should be performed to see under which
conditions the environment is stable.
Languages and/or other Software: Any, although investigation in the right type of language may pay off
Prerequisite:
none
Course Suitability: Knowledge in AI
Brief Description: Planning has been studied in AI since the
1970's. There are also approaches to extend it to automated program
construction. The idea is to use planning techniques and to generate
a program which satisfies a Hoare tripe {A}P{B}. The idea in this
project would be to use planning techniques and generate simple
programs which satisfy the specification. Unlike other approaches the
program has to be guaranteed to do what it is supposed to do.)
Languages and/or other Software: e.g., Lisp, Prolog, ...
Prerequisite:
none
Course Suitability: Knowledge in machine learning useful
Brief Description: The idea to this project is initiated by a talk in
the departmental seminar by Edmund Furse, there is a corresponding
paper with the title "A model of imitation learning of algorithms
from worked examples". The main idea is to present to a computer a
couple of examples and the computer is then able to generalise to a
general program. In this project, related ideas should be developed
and applied in a restricted application domain. A possible application
domain could be the generation of Emacs macros from examples. (The
editor Emacs has a macro-facility that allows to program a macro by
showing a single example. In all further examples the steps from the
first example are exactly taken over, in particular no generalisation
takes place. In this project the macro definition facility should be
extended so that it is possible to show different examples and the
macro created that way is a generalisation of all examples
presented.)
Languages and/or other Software: e.g., Pop11, Emacs Lisp
Prerequisite:
none
Literature: [Edmund Furse] A model of
imitation learning of algorithms from worked examples.
Course Suitability: Any
Brief Description: There is an approach to take input texts and
generate output texts which are generated from the input text by a
random selection that is biased by the input text. In this project a
similar approach should be followed for music. An interface for
comfortable input and output would be adequate. In order to generate
music at a more conceptional level, it might be useful to perform some
analysis of the music piece by compression algorithms as well.
Languages and/or other Software: Any
Prerequisite:
very good knowledge of music necessary, some
knowledge about machine learning useful
Course Suitability: AI
Brief Description: Many heuristics are subject to the so-called
no-free-lunch theorem, that is, they are good in some parts of the
search space but bad in others. Other heuristics are beneficial over
the whole class of all problems (of a particular kind). Typically
tests are difficult to perform since the class of all problems (of a
particular kind) may be very big (even infinite). The task is to come
up with a testing framework for heuristics, be they based on sampling
or testing the full class for small parameters.
Languages and/or other Software: Any, preferably an AI language
Prerequisite:
good knowledge in logic useful
Course Suitability: any, knowledge in logic required
Brief Description: Roughly speaking a model generator is a system that tries to generate
a finite model for a given set of first-order formulae. This is
essentially done by systematically searching in the space of possible
models, given by universes and interpretations. At first the system
tries whether there is a one-element universe, if not whether there is
a two-element one and so on. In order to be efficient in a huge search
space, choices have to be made in an intelligent way in order to
recognise dead ends as early as possible. Furthermore possibly
existing symmetries should be considered in order to minimise the
search effort. In this project a model generator is to be designed and
implemented.
Languages and/or other Software: Any
Prerequisite:
good knowledge in first-order logic
Literature: John Slaney.
FINDER: Finite domain enumerator-system description.
In Alan Bundy, ed., Proc. of the 12th CADE, Nancy 1994,
p. 798-801. Springer Verlag.
Course Suitability: any, knowledge in logic required
Brief Description: In this project should be tested whether proofs can be found by mining and composing them from existing proofs of related theorems. Typically humans learn how to find proofs by solving many related problems. Is it possible to do this on a shallow level of understanding and generate something which looks like a proof and then filter of all proof attempts those which are proofs? If you want to find out then this is the right project for you.
Prerequisite:
interest in web search, knowledge of first-order logic
Course Suitability: any, in particular AI
Brief Description: The development of classical logic was at least partially motivated by
the attempt to model aspects of human reasoning. However, classical
logic is not adequate to model human deduction (and is surely even
less appropriate for modelling other reasoning modes). One of the
problems is that in classical logic quite common and generally
accepted schemata of reasoning like modus ponens (from "A" and
"A implies B" it is possible to conclude "B") are
equivalent to less familiar one (like from "not B" and
"A implies B" it is possible to conclude "not
A"). However, applying the latter one causes much more
problems to human beings and an argument built up on the second form
is much likely to be faulty than one built up on modus ponens.
Such effects have been studied in detail by Braine et al. [Braine78], [Braine,Reiser,Rumain84] and Johnson-Laird, Byrne [Johnson-Laird,Byrne91].
In the proposed project the most recent literature has to be found and
studied, then it is aimed to develop from the cognitive model of the
reasoning process found there a prototype that can reproduce some of
the phenomena. In particular it would be interesting to investigate
the increase of error frequentness of human deduction under time
pressure.
Languages and/or other Software: CommonLisp, Pop11, or Prolog are particular well-suited
Prerequisite:
Some knowledge about classical logic is useful.
Literature: [Braine78]
Martin D.S. Braine.
On the relation between the natural logic of reasoning and standard logic.
Psychological Review, 85(1):1-21, 1978.
[Braine,Reiser,Rumain84]
Martin D.S. Braine, B.J. Reiser, and B. Rumain.
Some empirical justification for a theory of natural propositional logic.
Psychological Learning Motivation,
18(1):313-371, 1984.
[Johnson-Laird,Byrne 91] Philip Nicholas
Johnson-Laird and Ruth M.J. Byrne.
Deduction. Lawrence Erlbaum Associates Ltd., Hove, UK, 1991.
Course Suitability: any
Brief Description: Matching problems and algorithms which can
solve them are studied in economics to solve problems such as
matching schools and pupils, kidney donors and receivers, or
open job placements and applicants. Often legal rules must be
observed and it is necessary to be in the position to argue
that these rules are followed. In this project you would study
existing algorithms, implement your own, and reason on them.
Languages and/or other Software: Any programming language.
Prerequisite:
interest in economics
