Optimal Experimental Design for Uncertain Systems Based on Coupled Differential Equations

Youngjoon Hong, Bongsuk Kwon, Byung-Jun Yoon

Abstract

We consider the optimal experimental design problem for an uncertain Kuramoto model, which consists of N interacting oscillators described by coupled ordinary differential equations. The objective is to design experiments that can effectively reduce the uncertainty present in the coupling strengths between the oscillators, thereby minimizing the cost of robust control of the uncertain Kuramoto model. We demonstrate the importance of quantifying the operational impact of the potential experiments in designing optimal experiments.

Introduction

In scientific and engineering disciplines ranging from climate modeling to neuroscience, complex systems are frequently represented by coupled ordinary differential equations (ODEs). A persistent challenge in these fields is that the parameters governing interactions—such as coupling strengths—are often unknown or only partially characterized. When attempting to control such systems (e.g., to achieve synchronization), this uncertainty necessitates the use of "robust" control strategies. These strategies must account for the worst-case scenarios within the parameter uncertainty bounds, leading to significantly higher operational costs.

Traditional Optimal Experimental Design (OED) often focuses on reducing overall parameter variance or maximizing information gain (entropy). However, not all uncertainty is equally detrimental to the system's operational goals. This paper introduces an objective-based OED strategy that prioritizes experiments based on their ability to reduce the specific uncertainties that most heavily impact control costs.

Methodology

The Kuramoto Model and Control Objective

The authors utilize the Kuramoto model, a standard framework for studying synchronization in networks of $N$ oscillators. While the natural frequencies of the oscillators are assumed known, the pairwise coupling strengths between them are uncertain, defined only by upper and lower bounds. To achieve synchronization, an external controller oscillator is introduced. The objective is to minimize the interaction strength required by this controller to force the network into a synchronized state.

Mean Objective Cost of Uncertainty (MOCU)

The core of the proposed strategy is the Mean Objective Cost of Uncertainty (MOCU). MOCU quantifies the expected additional control cost incurred due to the lack of precise knowledge about system parameters. By minimizing MOCU, the experimental design process directly targets the reduction of operationally relevant uncertainty rather than statistical uncertainty alone.

Experimental Design and Physics-Informed Updates

The experimental space consists of pairwise synchronization trials. By isolating two oscillators, researchers can observe whether they synchronize. The authors prove a critical synchronization condition (Theorem 1): two isolated oscillators will synchronize if and only if the difference in their natural frequencies is less than or equal to twice their coupling strength. This binary outcome (sync or no-sync) allows for a rigorous update of the parameter bounds.

The OED algorithm simulates all possible pairwise experiments and selects the one that minimizes the expected remaining MOCU. To handle the computational intensity of Monte Carlo simulations involving ODE solvers, the authors implemented GPU acceleration using PyCUDA, reducing simulation times for 20,000 samples to under three seconds.

Results

The MOCU-based selection strategy was evaluated against two baselines: random selection and entropy-based selection (which targets parameters with the widest uncertainty).

Small-Scale Network (N=5)

In a system with 5 oscillators and 10 possible pairwise experiments, the MOCU-based approach demonstrated immediate effectiveness. As shown in Figure 1, it identified the most critical experiments within the first few updates, reaching the minimum possible uncertainty level in just 3 steps. In contrast, the entropy-based and random methods required significantly more updates to achieve similar performance.

MOCU reduction performance for a 5-oscillator system.

Medium-Scale Network (N=9)

The advantages of the MOCU-based approach were even more pronounced in a 9-oscillator system with 36 possible experiments. As illustrated in Figure 2, the MOCU-based method achieved a drastic reduction in the objective cost on the very first update. It reached the minimum uncertainty floor in 9 updates, while the entropy-based method stalled at a higher cost plateau for several iterations, and the random method lagged significantly behind.

MOCU reduction performance for a 9-oscillator system.

Conclusion

This research highlights that objective-based experimental design is superior to traditional variance-reduction techniques for controlling uncertain systems. By bridging the gap between statistical uncertainty and operational cost through MOCU, the proposed framework enables more efficient resource allocation in experimental science.

Future work will focus on extending this framework to handle noisier experimental outcomes and indirect observations through Bayesian inversion. Potential real-world applications include optimizing neuroimaging-based investigations of brain networks and refining noninvasive brain stimulation therapies for neurological disorders.

Summary updated

Abstract

We consider the optimal experimental design problem for an uncertain Kuramoto model, which consists of N interacting oscillators described by coupled ordinary differential equations. The objective is to design experiments that can effectively reduce the uncertainty present in the coupling strengths between the oscillators, thereby minimizing the cost of robust control of the uncertain Kuramoto model. We demonstrate the importance of quantifying the operational impact of the potential experiments in designing optimal experiments.
BibTeX
@article{hong2021optimal,
  author = {Youngjoon Hong and Bongsuk Kwon and Byung-Jun Yoon},
  title = {{Optimal Experimental Design for Uncertain Systems Based on Coupled Differential Equations}},
  journal = {IEEE Access},
  year = {2021},
  doi = {10.1109/access.2021.3071038},
}
Figure