Home

Home
(Japanese)

Concepts

Workshop

Groups

Members

Papers

Mail to:
cnb@atr.jp

�@

�@

Computational algorithms for acquisition of goal-directed behaviors have been developed in the theory of reinforcement learning. It has been shown in simulated learning experiments that the settings of several global parameters are critical for successful learning and that they determine the characters of learned behaviors. Based on the computational theory and neurobiological data on neuromodulators, we propose a set of hypotheses that different neuromodulators are used for active regulation of global parameters of learning in the brain.

1) The dopaminergic system encodes the increase or decrease in the cumulative future rewards and controls action selection.
2) The serotonergic system regulates the time scale of evaluation, i.e., how far into the future rewards are taken into account.
3) The noradrenergic system controls the 'temperature,' the balance between local optimization and global exploration.
4) The acetylcholinergic system controls the rate of learning and forgetting.

Based on these hypotheses, it is possible to predict how such global parameters, or the activities of neuromodulators, depend on each other and should interact with environments and learning processes.

(Possible Roles of Neuromodulators in the Regulation of Learning Processes, presented by Kenji Doya at the 30th Annual Meeting, Society for Neuroscience, November, 2000)