The core focus of our work is on investigating how we make decisions between options on the basis of their economic value (value-guided choice) and how we learn from the outcomes of our actions (reinforcement learning).

    We mostly use noninvasive neuroimaging techniques such as fMRI, MEG and EEG in task-performing healthy humans. A strong interest of the group lies in elucidating the roles of neurochemical systems in the behaviours of interest. Neuromodulators like dopamine have long been known to play a key role in learning and choice. However, while detailed knowledge is available on precise cellular actions of these neurotransmitters, comparably little is known about how they affect large-scale neural network dynamics. We start from the assumption that these neuromodulators guide behaviour by affecting larger ensemble network dynamics such as cortical oscillations, which we can record noninvasively in humans with MEG or EEG. We can either alter neurochemical activity by delivering drugs that target a certain system (e.g. dopamine receptor agonists or antagonists) or we can measure the concentrations of major neurotransmitter systems (GABA and glutamate) with MR spectroscopy. Furthermore, we often combine this with computational modelling approaches in order to get a window on the underlying computations of interest. 

                            

Jocham G, Hunt LT, Near J, Behrens TEJ. (2012) A mechanism for value-guided choice based on the excitation-inhibition balance in prefrontal cortex. Nat Neurosci; 15: 960–961.

     A second focus is on the neural mechanisms of model-based and hierarchical reinforcement learning. The classic models of reinforcement learning are „model-free“, meaning that they assume a simple learning from trial and error that is agnostic to any structure in the environment. However, humans often use such knowledge about the world to guide their choices. Imagine you work in a large building that you know rather well. One day, you discover a new food vending place in a certain area. Now, you may not have taken the shortest route to this place, instead you had to stop over at two colleagues‘ offices that are quite far away. A model-free learning system would now simply reinforce the entire sequence of actions that have lead you to the food place. Most likely, however, next time you want to get there, you’d exploit your knowledge of the building and simply try to take the shortest route to your rewarding food vendor. This is termed model-based learning.

    Another approach to augment learning is by setting subgoals and assessing each action with respect to whether the corresponding subgoal was achieved. When you are baking a cake and you notice that the dough is as thin as water, you will not proceed to the very end to find out that the cake does not look the way it should. Also, you will not need to wonder whether it is really the final icing on the cake that spoiled it. Instead, it will be clear that one of the dough preparation steps has to be credited. This is called hierarchical reinforcement learning. We will investigate, mostly using fMRI, how these approaches are realized in the human brain.

                  

Behrens TEJ, Jocham G. (2011) How to perfect a chocolate soufflé and other important problems. Neuron; 71: 203–205.          www.Southparkstudios.com