Research map / Learning theory and algorithms / Combinatorial and search optimization
FERERO: A Flexible Framework for Preference-Guided Multi-Objective Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Casts preference-guided multi-objective learning as constrained vector optimization (FERERO) with polyhedral-cone relative preferences and linear absolute constraints, and gives single-loop and stochastic primal algorithms with non-asymptotic convergence.
Problem
Finding specific preference-guided Pareto solutions reflecting different trade-offs among multiple objectives; existing methods are restrictive in preference definitions and/or lack theoretical guarantees.
Approach
Unifies relative preferences (partial order induced by a polyhedral cone C_A) and absolute preferences (linear constraints on objectives) in one constrained vector optimization problem, solved by an adaptive subprogram that needs no feasible start and no active-set identification, yielding a single-loop primal algorithm (FERERO-SA) with stochastic variant.
Key results
- On Multi-MNIST loss, FERERO achieves hypervolume 1.97 (x10^-2) vs 1.68 for LS and 1.42 for XWC-MGDA.
- On multilingual speech recognition (Librispeech + AISHELL v1), FERERO-FT reaches average WER 5.1% vs 6.4% for LS-FT.
- On the emotion dataset, FERERO's hypervolume (0.0357) is below EPO (0.0366), so it is not best there.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FERERO instead of LS | LeNet multi-task image classification · Multi-MNIST loss | hypervolume: 1.68 -> 1.97{ $\pm$0.21} | Helped |
| Use FERERO instead of XWC-MGDA | LeNet multi-task image classification · Multi-Fashion loss | hypervolume: 6.77 -> 7.76{ $\pm$0.18} | Helped |
| Use FERERO instead of XWC-MGDA | LeNet multi-task image classification · Multi-F+M loss | hypervolume: 3.89 -> 3.82{ $\pm$0.21} | Hurt |
| Use FERERO instead of LS | LeNet multi-task image classification · Multi-MNIST accuracy | hypervolume: 0.19 -> 0.24{ $\pm$0.04} | Helped |
| Use FERERO instead of LS | LeNet multi-task image classification · Multi-Fashion accuracy | hypervolume: 0.99 -> 1.17{ $\pm$0.07} | Helped |
| Use FERERO instead of XWC-MGDA | LeNet multi-task image classification · Multi-F+M accuracy | hypervolume: 0.52 -> 0.53{ $\pm$0.04} | Helped |
| Use FERERO instead of EPO | Emotion recognition · Emotion loss | hypervolume: 0.0366 -> 0.0357{ $\pm$0.0006} | Hurt |
| Finetune with FERERO instead of linear scalarization | M2ASR multilingual speech recognition · Librispeech (English) | WER (%): 6.8 -> 5.4 | Helped |
| Finetune with FERERO instead of linear scalarization | M2ASR multilingual speech recognition · AISHELL v1 (Chinese) | WER (%): 5.9 -> 4.9 | Helped |
| Finetune with FERERO instead of linear scalarization | M2ASR multilingual speech recognition · Average (Librispeech and AISHELL v1) | WER (%): 6.4 -> 5.1 | Helped |
| Include CPC loss in addition to supervised CTC loss | Multilingual speech recognition · Average WER | WER (%): +4.2% | Helped |
| Finetune with linear scalarization | Multilingual speech recognition with CPC loss · Average WER | WER (%): +0.3% | Helped |
Surprising findings
- On the Emotion loss hypervolume, EPO and XWC-MGDA outperform FERERO, and on Multi-F+M loss XWC-MGDA is best.
Limitations
- Higher per-iteration complexity than scalarization methods.
- Assumptions (bounded objectives, Lipschitz and smooth) may be restrictive; the PL-type assumption for sharper rate is not verified in general.
- Results do not uniformly beat baselines (e.g., Emotion and Multi-F+M losses).
Experimental setup
Datasets: Multi-MNIST Multi-Fashion Multi-Fashion+MNIST Emotions and Music Librispeech AISHELL v1
Benchmarks: Multi-MNIST Multi-Fashion Multi-F+M Emotions and Music Librispeech AISHELL v1 synthetic bi-objective problem
Metrics: hypervolume accuracy loss word error rate (WER)
Baselines: Linear scalarization MGDA PMTL EPO XWC-MGDA Smooth Tchebycheff
Builds on: MGDA EPO PMTL XWC-MGDA sequential quadratic programming
Closest papers by meaning
- Preference-Aware Constrained Multi-Objective Bayesian Optimization (Student Abstract) (AAAI 2024 · similarity 0.89)
- Preference-Guided Diffusion for Multi-Objective Offline Optimization (NeurIPS 2025 · similarity 0.89)
- Preference Optimization on Pareto Sets: On a Theory of Multi-Objective Optimization (NeurIPS 2025 · similarity 0.89)
- Multi-Objective Bayesian Optimization with Active Preference Learning (AAAI 2024 · similarity 0.89)
- User Preference Meets Pareto-Optimality in Multi-Objective Bayesian Optimization (AAAI 2025 · similarity 0.87)
- Pareto Merging: Multi-Objective Optimization for Preference-Aware Model Merging (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.