BUSTLE: A Versatile Tool for the Evolutionary Learning of STL Specifications from Data.

Area of Science:

• Complex Systems Analysis
• Machine Learning
• Formal Methods

Background:

• Monitoring and understanding complex systems requires describing their temporal properties.
• Signal Temporal Logic (STL) offers an expressive, human-readable framework for specifying system properties.
• Automatic learning of STL formulae from observational data is an emerging research area.

Purpose of the Study:

• To propose BUSTLE (Bi-level Universal STL Evolver), an evolutionary computation approach for learning STL formulae from data.
• To address limitations of existing methods by handling broader classes of system observation scenarios.
• To automatically learn both the structure and parameters of STL formulae.

Main Methods:

• Employs evolutionary computation with a bi-level search mechanism: global search for formula structure and local search for parameter values.
• Handles two distinct data availability cases: (a) observations of both regular and anomalous system states, and (b) only regular state observations.
• Experimental evaluation and comparison against prior approaches on relevant problem instances.

Main Results:

• BUSTLE successfully evolves effective and human-readable STL formulae for complex system monitoring.
• Demonstrates applicability across different data availability scenarios (regular/anomalous states vs. only regular states).
• Achieves comparable or improved effectiveness without sacrificing the human-readability of the learned formulae.

Conclusions:

• BUSTLE represents a significant advancement in the automatic learning of STL formulae from system data.
• The bi-level evolutionary approach enhances versatility, enabling application to a wider range of monitoring problems.
• The method provides a powerful tool for complex system analysis, balancing effectiveness with interpretability.