Neuron-level Prediction and Noise can Implement Flexible Reward-Seeking Behavior.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Traditional reinforcement learning often relies on explicit environmental reward functions.
• Autonomous agents require mechanisms for adaptive behavior and decision-making.
• Understanding biologically plausible learning rules is crucial for advancing AI.

Purpose of the Study:

• To demonstrate neural networks can achieve reward-seeking behavior without external rewards.
• To investigate the role of internal noise and local updates in autonomous behavior.
• To explore how networks adapt to environmental and architectural changes.

Main Methods:

• Development of neural networks utilizing local predictive updates and internal noise.
• Analysis of attractor dynamics governing explore-exploit switching.
• Testing network adaptability to modifications in architecture, environment, and motor interfaces.
• Investigating task preference formation and bias mechanisms.

Main Results:

• Neural networks successfully implemented reward-seeking behavior autonomously.
• Internal noise and local updates were sufficient for adaptive interaction.
• Networks demonstrated plasticity, adapting to changes without external control.
• Task preferences were shown to be influenced by noise, initialization, and network architecture.

Conclusions:

• A novel, biologically plausible algorithm enables autonomous, adaptable interaction with environments.
• The approach removes the need for explicit environmental reward functions.
• This work offers a flexible framework for developing self-governed intelligent agents.