[Paper] Frenetic Cat-inspired Particle Optimization: a Markov state-switching hybrid swarm optimizer with application to cardiac digital twinning

Source: arXiv - 2604.15761v1

Overview

The paper introduces Frenetic Cat‑inspired Particle Optimization (FCPO), a new hybrid swarm optimizer designed to squeeze the most out of very limited evaluation budgets. By blending particle‑swarm dynamics with a Markov‑state controller that toggles between exploration and refinement on the fly, the authors demonstrate that FCPO can solve expensive black‑box problems—such as calibrating cardiac digital twins—faster than many state‑of‑the‑art algorithms while keeping solution quality competitive.

Key Contributions

• Markov‑state switching controller that dynamically selects exploration vs. exploitation operators during a single run.
• State‑conditioned bounded motion to keep particles inside sensible regions while still allowing aggressive moves when needed.
• Elite‑difference global jump operator that injects large, directed jumps to break out of stagnation.
• Eigen‑space guided local refinement that leverages the covariance of elite solutions for efficient fine‑tuning.
• Linear population‑size reduction that trims computational cost in later stages without sacrificing convergence.
• Extensive benchmarking on CEC‑2022 functions (10‑dim & 20‑dim) and a real‑world cardiac ventricular activation calibration task, showing up to 2.6× speed‑up over leading optimizers like CMA‑ES and L‑SHADE.

Methodology

FCPO starts with a conventional particle swarm: each particle has a position and velocity, and it updates based on its own best experience and the global best. What makes FCPO different is a lightweight Markov chain that lives alongside the swarm. The chain has a few discrete states (e.g., “explore”, “refine”, “jump”), and at each iteration it probabilistically transitions based on simple metrics such as improvement rate or diversity.

• Exploration state: particles move with a bounded random walk, encouraging coverage of the search space.
• Refinement state: the algorithm computes the covariance matrix of the current elite set, extracts its eigen‑vectors, and nudges particles along the most promising directions (similar to a low‑cost quasi‑Newton step).
• Jump state: an elite‑difference vector (difference between the best and a randomly chosen elite) is added to a particle, creating a large, directed leap that can escape local minima.

The population size is linearly decreased as the run progresses, which reduces the number of expensive objective evaluations in the later, fine‑tuning phase. All components are implemented in pure Python, keeping the codebase lightweight and easy to integrate into existing pipelines.

Results & Findings

Key take‑aways:

• Runtime efficiency – Across all ten benchmark cases FCPO averaged 0.183 s, the fastest among the compared algorithms.
• Accuracy trade‑off – On highly structured functions (F1–F3) CMA‑ES still edged out FCPO in raw objective value, but FCPO’s runtime advantage makes it attractive when evaluation cost dominates.
• Robustness on real data – In the cardiac digital twin calibration, FCPO consistently converged to physiologically plausible activation maps across multiple random seeds, confirming its reliability for expensive inverse problems.

Practical Implications

1. Fast hyper‑parameter tuning for costly simulations – Engineers can plug FCPO into pipelines that involve CFD, finite‑element, or electrophysiology simulations where each function call may take seconds to minutes.
2. Real‑time or near‑real‑time model calibration – The ability to hit a target fidelity within ~40 iterations opens the door for adaptive digital twins that update on‑the‑fly as new patient data streams in.
3. Lightweight integration – Since the reference implementation is pure Python with no heavy dependencies, developers can drop FCPO into existing scikit‑optimize, optuna, or custom Bayesian‑optimization loops without a steep learning curve.
4. Resource‑aware optimization – The built‑in population‑size decay means you can set a hard budget (e.g., “no more than 5 k evaluations”) and let FCPO automatically allocate effort between exploration and exploitation.

Limitations & Future Work

• Benchmark scope – The study focuses on CEC‑2022 synthetic functions and a single cardiac application; broader testing on other high‑dimensional, noisy, or constrained problems would strengthen claims.
• Parameter sensitivity – While the Markov controller is simple, its transition probabilities and state thresholds still require manual tuning; an auto‑tuning layer could make FCPO truly plug‑and‑play.
• Scalability to very high dimensions – Experiments stop at 20 dimensions; performance on 100‑dim or higher spaces (common in deep‑learning hyper‑parameter search) remains unknown.
• Hybridization with surrogate models – Future work could combine FCPO’s state‑switching with surrogate‑based evaluations (e.g., Gaussian processes) to further cut down on expensive objective calls.

If you’re building a system that must squeeze the most out of a tight evaluation budget—whether it’s a medical digital twin, a physics‑based simulator, or a large‑scale hyper‑parameter sweep—FCPO offers a compelling blend of speed, adaptability, and ease of integration.

Authors

• Jorge Sánchez
• Guadalupe García-Isla
• Sandra Perez-Herrero
• Beatriz Trenor
• Javier Saiz

Paper Information

• arXiv ID: 2604.15761v1
• Categories: cs.NE, math.OC
• Published: April 17, 2026
• PDF: Download PDF