Less can be More: Effects of a Forgetting Function on an AI-based Policy Capturing Tool Performance

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Machine Learning

Background:

• AI systems require adaptability for dynamic environments.
• Human forgetting offers adaptive advantages.
• Concept drift challenges AI in changing circumstances.

Purpose of the Study:

• Investigate integrating a human-inspired forgetting function into AI.
• Enhance AI adaptability and performance in dynamic settings.
• Address concept drift by prioritizing recent information.

Main Methods:

• Implemented a training window (hyperparameter) in the Cognitive Shadow (CS) AI system.
• CS learns and emulates human decision models.
• Tested AI performance in a rock-paper-scissors-like game over three sessions.

Main Results:

• The training window significantly improved AI prediction accuracy in sessions 2 and 3.
• Prioritizing recent data via the training window enhanced performance.
• The AI successfully predicted and countered human decisions.

Conclusions:

• Human-inspired forgetting mechanisms can improve AI adaptability.
• The training window is effective in mitigating concept drift.
• This approach has implications for future AI decision support systems.