Self-organizing neural networks integrating domain knowledge and reinforcement learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Integrating domain knowledge into AI systems can improve learning efficiency and reduce complexity.
• Traditional reinforcement learning (RL) struggles to incorporate domain knowledge directly due to structural incompatibilities and its exploratory nature.

Purpose of the Study:

• To demonstrate how self-organizing neural networks (SONNs) can integrate domain knowledge with RL.
• To develop strategies for effective utilization of domain knowledge within SONNs during RL.

Main Methods:

• Symbol-based domain knowledge is translated into numeric patterns for insertion into SONNs.
• Analysis of how SONNs utilize inserted domain knowledge during RL.
• Development of a vigilance adaptation and greedy exploitation strategy.

Main Results:

• SONNs can effectively integrate and utilize domain knowledge within an RL framework.
• The proposed strategy enhances the exploitation of domain knowledge while maintaining learning plasticity.
• Experimental validation in pursuit-evasion and minefield navigation domains.

Conclusions:

• Self-organizing neural networks offer a viable method for integrating domain knowledge into reinforcement learning.
• The developed strategies improve learning efficiency and reduce model complexity in RL tasks.
• This approach facilitates more effective and efficient AI learning systems.