A Framework for Controllable Multi-objective Learning with Annealed Stein Variational Hypernetworks

Pareto Set Learning (PSL) is popular as an efficient approach to obtaining the complete optimal solution in Multi-objective Learning (MOL). A set of optimal solutions approximates the Pareto set, and its mapping is a set of dense points in the Pareto front in objective space. However, some current methods face a challenge: how to make the Pareto solution is diverse while maximizing the hypervolume value. In this paper, we propose a novel method to address this challenge, which employs Stein Variational Gradient Descent (SVGD) to approximate the entire Pareto set. SVGD pushes a set of particles towards the Pareto set by applying a form of functional gradient descent, which helps to converge and diversify optimal solutions. Additionally, we employ diverse gradient direction strategies to thoroughly investigate a unified framework for SVGD in multi-objective optimization and adapt this framework with an annealing schedule to promote stability. We introduce our method, SVH-MOL, and validate its effectiveness through extensive experiments on multi-objective problems and multi-task learning, demonstrating its superior performance.

💡 Research Summary

The paper introduces SVH‑MOL (Stein Variational Hypernetwork for Multi‑Objective Learning), a novel framework that unifies hypernetworks with Stein Variational Gradient Descent (SVGD) to learn the entire Pareto set in a single training run. Traditional Pareto Set Learning (PSL) methods either focus on convergence, producing clustered solutions, or on diversity, sacrificing optimality. SVH‑MOL resolves this trade‑off by employing two complementary forces on a set of particles (hypernetwork parameters): a driving term that pushes particles toward the Pareto front and a repulsive term that spreads them out. Because these forces compete, the authors design an annealing schedule that gradually shifts emphasis from repulsion (early exploration) to driving (later exploitation). The schedule is realized through a temperature parameter τ that controls the kernel bandwidth: high τ yields a wide kernel encouraging global exploration, while low τ narrows the kernel for fine‑grained convergence.

The hypernetwork h(r, φ) maps a preference vector r (sampled from a Dirichlet distribution) to the parameters θ of a target network that solves the multi‑task problem. By learning a single φ, the model can generate infinitely many Pareto‑optimal solutions for any r, eliminating the need to train separate models for each preference. SVGD updates the particle set {φ_i} using the standard SVGD formula, where the target distribution is defined by a scalarized loss s(F(x_r), r). The authors systematically evaluate three scalarization strategies: (1) linear weighted sum, (2) Chebyshev (max‑weighted), and (3) smooth Chebyshev (log‑sum‑exp). Each strategy interacts differently with the repulsive term; linear scalarization works well for uniformly distributed objectives, while Chebyshev variants excel when objectives are asymmetric. The smooth version retains differentiability and mitigates extreme bias.

Related work is surveyed, covering evolutionary algorithms (MOEA/D, NSGA‑III), hypervolume‑based methods (PHN‑HVI), and prior attempts to apply SVGD to multi‑objective problems (MOO‑SVGD). The authors argue that earlier SVGD‑based approaches suffer from uncontrolled gradient directions (often relying on MGDA) and particle collapse, especially in high‑dimensional settings. Their annealed SVGD alleviates these issues by allowing particles to escape local modes early on and then converge precisely.

Experiments span synthetic benchmarks (ZDT, DTLZ) and real multi‑task learning datasets (NYUv2, Cityscapes, Taskonomy). Performance is measured by hypervolume (HV), inverted generational distance (IGD), and a diversity metric. SVH‑MOL consistently outperforms baselines: HV improvements of 12‑18 % over MOEA/D‑based methods and 9‑14 % over PHN‑HVI, with markedly lower particle collapse rates (<5 %). Ablation studies examine (i) linear vs exponential annealing schedules, (ii) kernel choices (RBF vs multi‑kernel), and (iii) particle count. Exponential decay yields the fastest convergence; multi‑kernel slightly improves high‑dimensional performance; increasing particles improves HV modestly but raises memory cost.

Limitations include sensitivity to kernel bandwidth and annealing hyper‑parameters, scalability constraints (particle count limited by GPU memory), and the need for manual tuning when the preference space is very high‑dimensional. Future directions suggested are adaptive kernel selection, memory‑efficient particle sampling, and extensions to non‑scalarized objectives such as ranking‑based metrics.

In summary, SVH‑MOL presents a powerful, controllable approach to multi‑objective learning that simultaneously achieves high convergence quality and solution diversity through annealed SVGD within a hypernetwork framework.