PhD studentships
Department of Statistics
University of Bristol
Two studentships are available to enable intelligent and enthusiastic
individuals to join the 5*A Statistics group at the University of
Bristol, working on the inter-disiplinary ALADDIN project
(http://www.aladdinproject.org). These studentships will provide home
(UK or EU) fees and a maintenance allowance. Applicants should have a
first or upper second class honours degree in an appropriate subject,
such as statistics, maths, computer science, electrical engineering or
economics. The studentships will commence in October 2006.
ALADDIN is a multi-million pound multi-disciplinary research project
funded by a BAE Systems and EPSRC strategic partnership. It involves a
number of leading research groups from Imperial College London,
University of Southampton, University of Bristol, and Oxford University.
The project is concerned with developing mechanisms, architectures, and
techniques to deal with the dynamic and uncertain nature of distributed
and decentralised intelligent systems. Disaster management is the chosen
application domain as the world faces an urgent need for better means to
deal with such situations where a number of actors have to coordinate
their activities when facing significant degrees of uncertainty and
where the context is very dynamic. Thus, within ALADDIN, we will deal
with a number of research themes ranging from information fusion,
through multi-agent learning and Bayesian networks, to multi-agent
system architectures.
Successful applicants for these studentships will research the
mathematics of multi-agent reinforcement learning, and applications of
these ideas to the ALADDIN project. Reinforcement learning is a process
where individuals involved in a task select actions, receive rewards,
and attempt to improve their strategy based only on this information.
In a multi-agent (game-theoretical) setting the rewards available to an
individual from selecting an action depend on the strategies employed by
the other players, even though the actions of the other players might
never be directly observed. Furthermore, all the agents adapt their
strategies simultaneously, and agents may be required to learn to play
non-deterministic strategies, so this is a significantly more difficult
problem than either single-agent learning or traditional learning in games.
Mathematically, the process can be described as a large inhomogeneous
Markov chain. The asymptotic behaviour of several reinforcement
learning algorithms has been characterised when learning to play a
normal form game, but many inmportant open questions remain. In particular
- Do these algorithms converge to Nash equilibrium in more complex
situations?
- Can we ensure fast convergence to a good set of strategies (instead of
asymptotic convergence to Nash equilibrium)?
- How can learners make most efficient use of the information available
to them?
- What is the best way to apply these ideas in practice?
Further details are available from Dr. David Leslie (Tel: 0117 954 5901,
Email: [log in to unmask], Web: http://www.maths.bris.ac.uk/~madsl).
|