Explaining Aha! moments in artificial agents through IKE-XAI: Implicit Knowledge Extraction for eXplainable AI.

Area of Science:

• Developmental robotics
• Explainable AI
• Cognitive modeling

Background:

• Children develop mental representations during learning, mirroring how Machine Learning (ML) algorithms create latent representations.
• The Tower of Hanoi (TOH) task is a benchmark for studying problem-solving and cognitive development.

Purpose of the Study:

• To investigate the knowledge construction of artificial agents during learning.
• To develop a methodology for extracting implicit knowledge from agents' behavior.
• To provide explainable insights into AI learning processes.

Main Methods:

• A 3-step methodology named Implicit Knowledge Extraction with eXplainable Artificial Intelligence (IKE-XAI).
• Utilizing a Q-learning agent for the TOH task, a recurrent neural network for knowledge encoding, and post-hoc rule extraction for finite state automata.
• Employing graph representations for visual explanations of agent behavior.

Main Results:

• The IKE-XAI approach successfully extracts implicit knowledge as an automaton.
• The method provides global explanations of an agent's knowledge evolution during learning.
• Researchers can identify the 'Aha!' moment when an agent's knowledge stabilizes.

Conclusions:

• IKE-XAI offers a novel way to understand and explain the learning process in artificial agents.
• The methodology bridges developmental robotics, cognitive modeling, and explainable AI.
• This approach facilitates the analysis of observational behavior in human-machine interaction contexts.