Neural Networks With Motivation.

Area of Science:

• Computational neuroscience
• Behavioral economics
• Artificial intelligence

Background:

• Decision-making in animals is influenced by internal motivational states.
• The mathematical understanding of motivational salience in decision-making is limited.

Purpose of the Study:

• To propose a reinforcement learning framework using neural networks to model optimal behavior under dynamic motivational states.
• To investigate how motivational salience affects decision-making and behavior adaptation.

Main Methods:

• Developed a reinforcement learning framework with neural networks implementing Q-learning and motivational salience.
• Utilized a hierarchical manager-agent system for an integrated reinforcement learning algorithm.
• Trained the model in a Pavlovian conditioning setting.

Main Results:

• Neural networks adapted behavior to shifting agent needs without synaptic adjustments, showing potential for addictive behaviors.
• The model successfully inferred motivational states and behavior in a hierarchical system.
• Simulated neuronal responses in the model aligned with recordings from the ventral pallidum.

Conclusions:

• Motivation enables Q-learning networks to rapidly adapt behavior based on dynamic needs and reward modulation.
• The framework offers insights into the algorithmic basis of motivation and brain's motivational dynamics.
• This approach aids in interpreting behavioral data by inferring motivational states.