A Reinforcement Learning Model Equipped with Sensors for Generating Perception Patterns: Implementation of a...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Reinforcement Learning

Background:

• Existing reinforcement learning (RL) techniques are often field-specific, limiting generalization and extrapolation.
• Current reward-punishment (r-p) learning processes and result convergence lack sufficient speed and efficiency.
• A need exists for a more generalized and efficient RL model applicable across various domains.

Purpose of the Study:

• To propose a general reinforcement learning model independent of input/output types.
• To enhance the speed and efficiency of the learning process using bio-inspired principles.
• To address the limitations of specialization and slow convergence in existing RL methods.

Main Methods:

• Developed a general RL model featuring a perception module that maps sensor inputs to generalized perception patterns.
• Implemented a statistical association procedure linking perception-action pattern pairs based on learning outcomes.
• Utilized a mechanism for rating results via positive/negative reactions to sensory stimuli and an adaptable action module.

Main Results:

• The proposed model was successfully applied in the air navigation domain, a highly safety-critical field.
• A simulated system equipped with the model used Automatic Dependent Surveillance-Broadcast (ADS-B) technology for perception.
• The model demonstrated superior performance compared to traditional methods in terms of learning reliability and efficiency.

Conclusions:

• The developed general reinforcement learning model offers improved generalization and faster learning.
• Bio-inspired principles effectively accelerate the learning process and enhance model adaptability.
• The model's successful application in air navigation validates its potential for complex, safety-restricted domains.