Reader in Computing at City University, UK.
Dates of visit: 02/10/2012 - 14/12/2012
Project Description: My research focuses on developing mathematical and computational models of learning and behavior as well as in studying the uses, abuses and misuses of such models in building psychological theories. More specifically, I am interested in applying machine learning and control algorithms to model decision-making processes. We have proposed a reinforcement learning method, value-gradient learning (VGL), that under certain conditions finds the optimal policy, the one that maximizes a cumulative reward. Unlike traditional approaches such as Temporal Difference methods where agents are model-free, VGL works with full models that are used in the defining the gradient of the values of states (or of state-action pairs) rather than mere values. In so doing, VGL reduces drastically the need for stochastic exploration and guarantees optimality in continuous state spaces. Our current research aims at selecting models for our learning machinery, in particular economic, social and political models -hence our interest in collaborating with The LSE Choice Group. In short, our theory serves as a bridge between Adaptive Dynamic Programming and from Reinforcement Learning in proposing hybrid mechanisms that respond to the new challenges as posed by research in neuroeconomics. Relatedly, VGL is equivalent to a variational principle, Pontryagin's Minimum Principle -in turn a version of the Hamilton's Principle of Least Action. This relation allows us to study cognition on the same bases as any other natural phenomena. Our intuition is that the dynamic laws derived from such principle will allow us to define conserved quantities and symmetries in psychology. With this aim in mind we would like to collaborate with the Philosophy of Physics project. Firstly, we think that the philosophical analysis of the concept of symmetry and symmetry breaking in information processing can be applied to physical and psychological systems alike; secondly, the results of such analysis will be reflected in a formal model specified in the groups (and possibly categories) that generate the symmetries.