Daniel Braun

One of the most striking features that sets human motor control apart from its robotic counterparts is the remarkable adaptability that allows us to cope with a vast range of complex and variable environments. The research goal of our group is to investigate the computational and biological principles underlying this unrivalled adaptability both experimentally and theoretically. In behavioural experiments, virtual reality technology allows us to expose human subjects to diverse, novel and possibly complex environments. Our aim is to study how the human motor system exploits the structure and the causal dependencies in such environments to enhance adaptation and to integrate information for action. These experimental studies are backed up by theoretical investigations of normative principles for adaptive control to explain the observed behaviours under special consideration of the bounded resources of the actors.

The interests of the research group can be broken down into three sub-projects:

(A) Structural learning and causal inference in motor control

In our previous studies we have provided first evidence for structural learning processes in human motor control, as we could show that the motor system extracts (abstract) invariants when faced with a range of variable environments that share certain structural features. However, many questions remain open that still need to be addressed, including the basic experimental protocol. In particular, more quantitative experiments are needed to understand the computational principles behind structural motor learning. This also requires the development of computational models that allow for quantitative predictions of structural learning and causal inference in human motor control.

(B) Neuro-economical principles in motor control and learning

In our previous studies we have investigated the effects of risk-sensitivity and two-player interaction on motor control. We could show that risk-sensitive subjects are not only evaluating the expected value of a movement, but also consider second-order moments of the uncertainty and act either optimistically (risk-seeking) or pessimistically (risk-aversive) with respect to this risk. In a set of two-player motor interaction experiments, we could show that subjects’ motor behavior naturally converged to so-called Nash solutions that describe rational strategies in game theory. Both risk and the presence of another actor are important neuroeconomic constraints both experimentally and in normative models of acting. However, important open questions remain about how uncertainty over different structures ("model uncertainty") is integrated by human subjects, how sensory uncertainty is affected by risk-attitudes and how uncertainty over different states of information in two-player interactions affects sensorimotor behaviour.

(C) Resource-bounded principles of adaption and control

In our previous studies we have developed a Bayesian rule for adaptive control based on an information-theoretic compression principle and causal calculus. We have demonstrated the applicability of the new adaptive control framework for very general problem classes such as undiscounted Markov Decision Problems, adaptive linear quadratic control problems and bandit problems. Important open problems are applications to more complex environments requiring generic non-parametric controllers, and in particular the application to environments that are themselves adaptive, which leads to game-theoretic problems. We have also worked on a generalization of the information-theoretic compression principle to describe normative behavior in agents with bounded computational resources. We explore further in how far this compression principle can be exploited in the development of efficient algorithms for learning and control.