Stabilizing Bandit Mixture-of-Experts and Zipfian Stream Clustering via Router Replay and Clamped Laguerre Potentials
Year: 2026
Type: preprint
Abstract: Discrete routing rules assign specialist experts, cluster prototypes, or memory blocks without evaluating unchosen candidates. In partial-feedback or imbalanced-stream settings, standard gradient and optimal-transport updates lead to unstable assignments. When routing parameters are updated across multiple epochs on logged bandit data, router weights drift from the behavior distribution. This drift flips expert assignments and distorts credit attribution. Similarly, in online clustering with bursty Zipfian class arrivals, prototype collapse occurs when optimal-transport potentials enforce uniform marginals within micro-batches or accumulate unbounded load debt. Here, we propose two methods: Rollout Router Replay with Dense Control-Variate Mixing (R3-DenseMixer) for contextual-bandit mixtures of experts, and Windowed Clamped Laguerre Debt for imbalanced stream clustering. Router replay fixes routing decisions to the behavior rollout during inner updates and centers gradients around an all-expert critic, giving unbiased updates with lower variance. Clamped Laguerre potentials restrict load debt over a sliding window inside dispersion-scaled power diagrams, ensuring invariance to positive scale in each coordinate and preserving rare prototypes. On 30 unseen contextual-bandit benchmarks, R3-DenseMixer achieves a median test loss of 0.6928 (p_holm <= 0.00736 against all baselines). In bursty Zipfian streams, clamped Laguerre routing raises median Adjusted Rand Index to 0.6024.
Author: Coutinho, Felipe Provezano
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7236758241
