Multiple representations and algorithms for reinforcement learning in the cortico-basal ganglia circuit

Area of Science:

• Neuroscience
• Computational Neuroscience
• Cognitive Science

Background:

• The neural network of the cerebral cortex and basal ganglia is crucial for reinforcement learning.
• Recent research highlights functional heterogeneity within the cortico-basal ganglia circuit along its ventromedial to dorsolateral axis.

Purpose of the Study:

• To review computational issues in reinforcement learning.
• To propose a hypothesis on how multiple reinforcement learning algorithms are implemented in the cortico-basal ganglia circuit.

Main Methods:

• Review of existing literature on reinforcement learning and cortico-basal ganglia function.
• Development of a working hypothesis integrating computational principles with neurobiological evidence.

Main Results:

• Identification of key computational challenges in reinforcement learning.
• Proposal of a framework for understanding the implementation of diverse reinforcement learning algorithms within the brain.
• Emphasis on the role of different representations of states, values, and actions.

Conclusions:

• The cortico-basal ganglia circuit likely employs multiple, distinct reinforcement learning algorithms.
• Functional specialization along the ventromedial to dorsolateral axis supports this algorithmic diversity.
• Understanding these mechanisms offers insights into learning, decision-making, and neurological disorders.