A nonlinear hidden layer enables actor-critic agents to learn multiple paired association navigation.

Area of Science:

• Neuroscience and computational modeling
• Rodent learning and navigation

Background:

• Multiple cued reward location tasks are vital for studying rodent learning.
• Deep reinforcement learning agents can learn these tasks but lack biological plausibility.
• Classic biologically plausible actor-critic agents are limited to single reward locations.

Purpose of the Study:

• To investigate biologically plausible agents for multiple cue-reward location navigation.
• To identify limitations of classic actor-critic agents in complex navigation tasks.
• To develop and test novel computational models for enhanced rodent learning.

Main Methods:

• Computational modeling of actor-critic agents.
• Adaptation of agents to single reward locations and displacements.
• Introduction of a feedforward nonlinear hidden layer processing place cell and cue information.
• Implementation of temporal difference error-modulated plasticity.
• Comparison with a recurrent reservoir network architecture.

Main Results:

• Classic actor-critic agents navigated single reward locations but failed with multiple associations.
• A novel agent with a feedforward hidden layer successfully learned multiple cue-reward navigation.
• Replacing the feedforward layer with a recurrent reservoir network led to faster learning.

Conclusions:

• Classic biologically plausible agents are insufficient for complex multi-cue navigation tasks.
• A novel computational approach incorporating a hidden layer with TD error-modulated plasticity enhances agent capabilities.
• Recurrent networks offer further improvements in learning speed for these navigation tasks.