I am happy to supervise students on a number of topics related to models of perception, decision making and learning in humans. Below are two examples of possible projects.

(Dr. Ulrik Beierholm, email: u.beierholm@bham.ac.uk)

Computational advantage of multiple controllers in reinforcement learning

Reinforcement learning is a powerful way of modelling the learning of an organism acting in an environment with potential rewards and punishments. While it is well known that model based versions of RL are most effective, they very fast become computationally intractable as the number of options increase. In contrast a model-free system is computationally simple, but only converges slowly to the correct values. It has been proposed that a system that utilizes both models in parallel and which arbitrates between them based on their own uncertainty can work as a computationally plausible yet efficient solution.

This study will use numerical simulations to test for any computational advantage in using such multiple controller. The placement will be largely based on developing, programming and numerically testing a model of multiple controllers for RL

Non-parametric statistical models in human learning and perception

Our understanding of how humans learn about our environments, and how this relates to our perception of objects and structures in our environment, has been expended in the last 15 years through the use of Bayesian statistical models. However so far this work has almost exclusively relied on parametric models (such as Gaussian) which although simple in structure, are limited in requiring a pre-specified complexity (e.g. In mixture of Gaussian models). Meanwhile modern statistics and machine learning has developed a number of tools based on non-parametric statistics (such as Dirichlet processes) which allow the data itself to specify the complexity of the model; little data leads to simple models while a large dataset will naturally lead to a more complex model.

This project proposes to utilise such non-parametric models to further our understanding of both learning and perception in human cognition. The work will involve a combination of analytical model development, numerical simulation and possibly human psychophysical experiments.