All known stochastic optimisation methods such as Simulated Annealing, Evolutionary Algorithms and Estimation of Distribution Algorithms have a range of adjustable parameters like learning rates, crossover probabilities and weighting factors. Poor algorithm parameterisation hinders the discovery of good solutions. We are investigating ways of adjusting and adapting algorithm parameters during the optimisation process, with the goal of achieving optimal algorithm performance.
Problem difficulty is the search cost or computational complexity that results from the interaction of the search algorithm with the search space. This interaction can be modelled using the concept of fitness landscape, which relates the topological description of the search space with the dynamics of the search algorithms. We use fitness landscapes to study the structure of a search space, and analyse features that make optimisation problems hard to solve.
Software testing is a crucial part of software development. It enables quality assurance, such as correctness, completeness and high reliability of the software systems. We have built a Kalman filter-based genetic algorithm for solving the problem of test case generation in software testing. The method uses a Kalman filter to reduce the effect of the stochastic behaviour of GAs when estimating the appropriate parameter values to use in each iteration of the optimisation process.
The design of embedded systems, and in particular of automotive embedded systems involves several important decisions, such as which software components to select and how to deploy them into the hardware architecture. These decisions affect different quality attributes of the software system, such as reliability and safety. With the increasing number of functions performed by software, embedded systems are becoming more complex with many design options to choose from. We have developed optimisation techniques to automate this task.
After the Big Bang it took about 300,000,000 yr before the first stars would form - now some 13,000,000,000 yr ago. Unfortunately, we can no longer observe these stars today directly, even with our best telescopes. But there is still some 'fossil' record of them left behind, preserved in the oldest stars in our galaxy we can observe, dating back to pre-galactic times. When the first stars exploded as supernovae, their ashes were dispersed and the next generation of stars formed. We can now measure these abundance patterns in those old stars. The aim is to find a match and combination of 'ashes' from theoretical models in a large database containing a wide variety of stellar models and supernova and compare to observational data.
Monte-Carlo tree search is a powerful game playing algorithm which uses stochastic methods to evaluate game states. It has performed extremely well in some games, but has proven less useful in others, notably chess. This deficiency prompted us to investigate the efficacy of MCTS in finding sacrifice moves, moves which require some resource to be deliberately given up in order to open up a good sequence of moves.