Conditioning and Reinforcement learning
Conditioning experiments have a long tradition in
biology and psychology. The basic pardigm can
be formulated as reward-based `reinforcement' learning.
From a mathematical point of view, reinforcement learning (Sutton and Barto)
is a class of machine learning algorithms
that can be understood as iterative solutions of
the Bellman equation (dynamic programing).
A basic element in the learning rule is a reinforcement
signal that is positive only if the actual reward
is larger than the expected reward.
The group of Wolfgang Schultz (previously in Fribourg, now
in Cambridge) has found
exactly this type of activity in dopaminergic
neurons in the basal ganglia.
The connection between basal ganglia signals
on the one side and conditioning and reinforcement
learning on the other side has been recognized for a long time.
In this project, we
adress, among others, the following questions:
-
What are the timing conditions in conditioning
in relation to basal ganglia signals?
- Can we systematically connect the biological
rules to formal theories of reinforcement learning
and optimization?
Collaborators:
Julien Mayor,
PhD student at the LCN.
This project (financed by the Swiss National Science
Foundation) is performed in collaboration with
the group of
W. Schultz in Cambridge/England.
Please send comments on this page to: [email protected]
Go to
LCN -
DI -
EPFL
LCN - Laboratory of Computational Neuroscience