Machine Learning-Based Stability Prediction for Magnet Packing in the Game Kluster
Abstract In the magnetic tabletop game Kluster, players sequentially place permanent magnets inside a cord-bounded arena without triggering inter-magnet collisions. Determining whether a candidate placement will relax into static equilibrium or snap into contact requires numerically integrating stiff, non-smooth equations of motion governed by steep 1/ r 4 dipole forces and dry Coulomb friction. This study investigates whether machine learning (ML) classifiers wrapped in size-conditional (Mondrian) Conformal Risk Control (CRC) can pre-screen candidate placements—discarding likely collisions prior to numerical integration—while bounding the rate of falsely discarded safe placements (the wrong-skip risk) below a user-specified tolerance α . We model identical cylindrical neodymium magnets as vertical point dipoles translating on a horizontal plane inside a rigid circular inelastic boundary. Using a Numba-compiled semi-implicit Euler simulator with step-halving convergence verification, we generate a dataset of 205,856 candidate placements across 1,500 sequential games with independently randomized friction coefficients and dipole force scales. Empirical evaluation across five experimental protocols reveals both the utility and the strict boundaries of surrogate screening: (1) Size-conditional CRC controls wrong-skip risk on average only across repeated game splits; for Histogram Gradient Boosting with force-ratio features at α = 0.05 in board-size bin 5–8, the mean wrong-skip rate is 0.0457 (SE = 0.0012), yet 33% of individual calibration-test splits exceed α due to finite-sample variance. (2) Paired comparisons against an analytical Combined physics rule show no statistically significant ML gain on sparse boards (bin 2–4: +0.0015, 95% CI [− 0.0010, + 0.0056]), but significant gains on crowded boards where multi-body forces dominate (bin 5–8: +0.0581; bin 9–14: +0.0379). (3) Under unadapted physical parameter shifts ( q < 0.8), empirical risk guarantees fail (0.0670 for gradient boosting vs. 0.1720 for static distance rules); explicit parameter conditioning restores overall risk control (0.0469) and central subgroup control (0.0459), though rare tail regimes remain above α due to small sample sizes. (4) A multi-layer perceptron neural network baseline requires more than twice the training time of gradient boosting (18.32 s vs. 8.61 s) without improving screening skip rates. On an AMD Ryzen 7 5800HS processor, hybrid ML screening reduces per-candidate latency from 6.78 ms (single run) and 20.075 ms (convergence-checked run) to 0.5709 ms for screened candidates, yielding net wall-clock speed-ups of 1.82× and 2.03×, respectively.
Authors
- Mayank Goswami (ORCID: https://orcid.org/0000-0001-8829-9980)
- Sankalp Savarn
Institutions
- Indian Institute of Technology Roorkee (IN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-17
- DOI
- https://doi.org/10.21203/rs.3.rs-11733631/v1
- Primary Topic
- Reinforcement Learning in Robotics
- Type
- preprint