Abstract
Techniques are disclosed for coordinating optimization across multi-layer recommendation pipelines in which each layer optimizes a proxy metric. For each layer, an optimization distance between a current policy and a reference policy is measured, for example using expected KL divergence, and constrained by a per-layer budget using training-time penalties and/or serving-time interpolation toward the reference policy. A total optimization pressure is computed from per-layer distances and cross-layer interaction terms weighted by sensitivity between layers, and per-layer budgets are allocated and reallocated under a total budget constraint. Satisficing thresholds cap proxy optimization after diminishing returns and may redirect optimization to secondary objectives. Budgets are adaptively recalibrated based on proxy-to-ground-truth correlation changes. A circuit breaker detects cascade onset across multiple layers and freezes updates, resets budgets, and rolls back toward reference policies. An API may expose budget status and headroom.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Optimization Budget Allocation System for Preventing Goodhart Breakpoint Crossing in Multi-Layer Recommendation Pipelines", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10738