JLS.1 Studies of Ant Colony Optimisation Algorithms (ACS, CS, CMIM, CaS)

The ant colony optimisation algorithm was developed by Dorigo about 10 years ago. It is motivated by how real ant colonies find the short routes around obstacles.
It has been widely applied to routing problems such as the travelling salesman problem, and route optimisation in telecommunication networks, and has recently been
applied to other fields. See Dorigo's Ant Colony Optimization Page at http://iridia.ulb.ac.be/~mdorigo/ACO/ACO.html for more information.

However, ant colony algorithms can be view in a wider context. They are distributed implementations of an algorithm which is related to other, well-studied search
algorithms. The ant colony algorithm basically has a population of agents (ants) which are each carrying out approximately greedy search. At the same time, the
agents are trying to learn a local value function on which greedy search produces near optimal solutions. As such, is is related both to reinforcement learning, and to
approximately-greedy search algorithms such as simulated annealing and the Go-With-the-Winner algorithm.

The goal of this project is to study the ant colony algorithm on search spaces which are constructed to have known properties and learn about its performance when
compared to these other algorithms.

JLS.2 A New Method for Learning a Mixture of Local Models (ACS,CaS, CMIM)

In a recent paper, Williams and Titsias (2003) developed a model for discovering object in images. This involved making a probabilistic model for
each possible object. Normally, such a model would suffer from combinatorial explosion, there being an exponential (in the number of objects)
number of possibilities to consider. They get around this by learning one object at a time, and removing them after they are found. In order to do this,
they need to use a robust approach, which is able to learn one object without contamination from other objects. To do this, they use a mixture model
with a "robustifying component" which is common for all objects and background.

The goal of this work is to apply this approach to supervised learning. Instead of learning about localized objects in an image, the system will learn
local rules which predict output variables from input variables. Each rule will be learnt one at a time, using a mixture model with a robustifying
component, and learning via an EM algorithm. Such an approach may be particularly useful in problems in which there are regions of predictability
against a background which is hard to predict (as is often supposed about the stock market). This project will attempt to prove the concept.

It would be very helpful if the person doing this project has taken the Machine Learning module, or the Neural Networks module.

Reference: Christopher K.I. Williams and Michalis K. Titsias. Learning About Multiple Objects in Images: Factorial Learning without Factorial
Search . To appear in Advances in NIPS 15, 2003.

JLS.3 Algorithmic Music Composition on Simulated Autonomous Instruments by the Learning of Statistical Similarity. (ACS, CS, CaS, CMIM)

A player piano is an autonomous musical instrument which plays an exact copy of some piece played by a human player. It allows no interaction. A
set of wind chimes in a gentle breeze can also be thought of an autonomous instrument. It differs from the player piano in that it has variation in what
it plays and can be influenced by a user, but it plays only very simple sounds.

The algorithmic composer Assaf Talmudi has proposed and developed a combination of the two. A fairly complex dynamical system is constructed (in
computer simulation) consisting of bells connected by springs. The properties of this system are constructed (using a search algorithm) to have
similar statistical properties of some musical piece. Thus, the instrument produces music which can be influenced by the user, but still has some
statistical similarity with a piece of music. The statistical property used in the work of Talmudi was the distribution of pitches.

The idea of this project is to extend this idea to more complex statistics, such as the distribution of time intervals, or the pairwise correlation between
notes. It may also be interesting to compose more complex pieces by creating several different instruments of this form, and use them together.

JLS.4 Machine Learning and Optimisation (ACS, CS, CaS, CMIM)

A new class of optimisation algorithms, called Estimation of Distribution algorithms (EDAs), uses machine learning to try to learn where good
solutions to a problem are. The method works like this: a probability model (usually a graphical model) generates test solutions to the problem. The
better solutions are selected, using some selection method. The probability model is then updated. so that the selected solutions are more likely to be
generated. This is repeated, in the hope that the probability distribution will evolve to one which generates good solutions with a high probability.

This approach is related to genetic algorithms, where mutation and crossover are replaced by sampling from the distribution. Sampling generates new
solutions which have similar statistics to the recently sampled solutions, just like mutation and crossover generate new strings which have similar
statistics to the "parent" strings. For a review on EDAs, see [1].

I have been studying these algorithms and have a number of studies I would like to see done. I will discuss possible specific projects in a meeting.

References: P. Larranaga and J. A. Lozano, "Estimation of Distribution Algorithms, A New Tool for Evolutionary Computation", Kluwer Academic
Publishers, 2001.

J. L. Shapiro, "The Sensitivity of PBIL to the Learning Rate, and How Detailed Balance Can Remove It", Foundations of Genetic Algorithms 7,
2003.