Abstract JanusMooCmixSampler is a plug-and-play, training-free multi-objective sampler that augments Optuna’s NSGA-II host with JANUS-style, geometry-guided candidate generation (arXiv:2608.22862).
Population optimizers such as NSGA-II drive search mainly through rank-based selection signals: a signal says one candidate dominates another, but not the local direction responsible for the improvement. JANUS recovers this missing local geometric signal without replacing the host optimizer. It estimates a local metric from the recent evaluation trace and occasionally replaces a child with inverse-curvature Gaussian samples drawn around the current Pareto front.