Ghost hunting in the nonlinear dynamic machine.

Area of Science:

• Dynamical Systems Theory
• Machine Learning
• Nonlinear Dynamics

Background:

• Analyzing complex dynamical systems often requires advanced computational methods.
• Interpreting the results of machine learning models applied to time-series data can be challenging.
• Dynamical systems theory offers a framework for understanding system behavior and stability.

Purpose of the Study:

• To develop and validate an integrated approach combining dynamical systems modeling and machine learning.
• To demonstrate the ability to recover temporal dynamics and characterize system topology from data.
• To classify distinct temporal patterns in complex systems, even without known governing equations.

Main Methods:

• Utilized random forest models to analyze simulated and real-world time-series data.
• Applied dynamical systems theory to interpret machine learning model outputs.
• Extracted set points (points of no change) and predicted changes to characterize system topology.
• Tested the approach on data simulated using a modified Cusp Catastrophe Monte Carlo and real-world accelerometer data.

Main Results:

• The integrated model successfully recovered the temporal dynamics and cusp catastrophe from simulated data, even with significant error variance.
• The approach differentiated movement dynamics patterns in accelerometer data, identifying set points for cyclic motion (walking) and attraction (stair climbing).
• The method provided predictions of future system states and decomposed the topological features implied by the temporal dynamics.

Conclusions:

• Integrating machine learning with dynamical systems modeling offers a powerful, exploratory nonlinear solution for analyzing dynamical systems data.
• This approach provides a pathway to interpret complex results and characterize system topology.
• The method is viable for classifying distinct temporal patterns in various systems, including those with unknown nonlinear dynamics.