A differential Hebbian framework for biologically-plausible motor control.

Area of Science:

• Neuroscience
• Machine Learning
• Control Theory

Background:

• Developing biologically plausible neural control architectures for autonomous learning is a significant challenge.
• Existing methods often struggle with systems where control signals produce non-monotonic error responses.

Purpose of the Study:

• To propose a novel neural control architecture capable of fully autonomous learning.
• To demonstrate its effectiveness in controlling systems with both monotonic and non-monotonic error responses.
• To explore the integration of feedback control with neural reinforcement learning.

Main Methods:

• Utilizing feedback controllers that learn desired states by selecting error signals.
• Employing a family of differential Hebbian learning rules for error-driven control.
• Coupling neural reinforcement learning with feedback control for non-monotonic error systems.

Main Results:

• The architecture successfully learns to control systems with monotonically responding errors.
• Integration with reinforcement learning effectively handles non-monotonic error responses.
• Feedback control simplifies the reinforcement learning problem by focusing on desired values.

Conclusions:

• The proposed neural control architecture offers a biologically plausible and effective approach to autonomous learning.
• This framework has the potential to simplify complex reinforcement learning tasks, enabling the learning of more sophisticated actions.
• The architecture can be extended to hierarchical systems for enhanced control capabilities.