A Layered Learning Approach to Scaling in Learning Classifier Systems for Boolean Problems

Area of Science:

• Artificial Intelligence
• Machine Learning
• Evolutionary Computation

Background:

• Evolutionary Computation (EC) typically discards learned knowledge between problem instances.
• Human learning effectively reuses knowledge from simpler problems for complex tasks.
• Existing Learning Classifier Systems (LCSs) show potential for knowledge reuse via Code Fragments but lack efficient mechanisms.

Purpose of the Study:

• To investigate how LCS can adopt a layered learning framework for efficient problem-solving.
• To develop an LCS capable of reusing learned knowledge across increasingly complex problems.
• To address the inefficiency of random knowledge reuse in LCS.

Main Methods:

• Development of XCSCF*, an LCS incorporating base axioms for learning and transfer learning.
• Recasting learning as a decomposition into subordinate problems, forming a curriculum.
• Training XCSCF* on a series of related problems to facilitate knowledge transfer.

Main Results:

• XCSCF* successfully captures general logic across diverse domains from a 'tabula rasa' state.
• The system efficiently solves various n-bit problems including Multiplexer, Carry-one, Majority-on, and Even-parity.
• Demonstrated effective reuse of learned knowledge in subsequent, more complex problems.

Conclusions:

• The developed XCSCF* framework facilitates efficient layered learning in LCS.
• This research represents a significant step towards achieving continual learning in artificial agents.
• Learned knowledge is effectively retained and reused, overcoming limitations of traditional EC approaches.