Research map / Learning theory and algorithms / Combinatorial and search optimization
Offline Multi-Objective Optimization
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.
Introduces Off-MOO-Bench, the first benchmark for offline multi-objective optimization spanning synthetic, NAS, RL, combinatorial, scientific and engineering tasks, and empirically adapts offline SOO and MOBO methods to it.
Problem
Offline multi-objective optimization lacks benchmarks and methods: only a fixed dataset is available, no online evaluation is allowed, and existing offline SOO benchmarks (Design-Bench) do not cover multiple conflicting objectives.
Approach
Builds a unified benchmark (Off-MOO-Bench) with offline datasets collected from MOEAs for diverse MOO tasks, and analyzes DNN-based (end-to-end, multi-head, multiple-model with offline SOO techniques) and GP-based MOBO offline methods within a common framework, using NSGA-II as the search step and hypervolume as the metric.
Key results
- Multiple Models + IOM has the best overall average rank (4.61), and Multiple Models (5.02) is runner-up, both beating the best training-set solution D(best) (10.03).
- No method beats D(best) on MOCO (permutation search spaces), showing discrete large-scale tasks remain hard.
- Best GP-based MOBO achieves average rank 8.64, MOBO-qParEGO 11.68 and MOBO-JES 10.30, all weaker than NN-based methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use multiple independent objective models with IOM instead of the best training-data set | Off-MOO-Bench offline MOO · Off-MOO-Bench average | average rank: 10.03 ± 0.07 -> 4.61 ± 0.05 | Helped |
| Use multiple independent objective models instead of an end-to-end model (ablation) | Off-MOO-Bench offline MOO · Off-MOO-Bench average | average rank: 7.32 ± 0.01 -> 5.02 ± 0.03 | Helped |
| Add IOM to multiple independent objective models (ablation) | Multiple Models · Off-MOO-Bench average | average rank: 5.02 ± 0.03 -> 4.61 ± 0.05 | Helped |
| Add COMs to multiple independent objective models (ablation) | Multiple Models · Off-MOO-Bench average | average rank: 5.02 ± 0.03 -> 8.09 ± 0.32 | Hurt |
| Add RoMA to multiple independent objective models (ablation) | Multiple Models · Off-MOO-Bench average | average rank: 5.02 ± 0.03 -> 7.07 ± 0.02 | Hurt |
| Add ICT to multiple independent objective models (ablation) | Multiple Models · Off-MOO-Bench average | average rank: 5.02 ± 0.03 -> 9.64 ± 0.11 | Hurt |
| Add Tri-Mentoring to multiple independent objective models (ablation) | Multiple Models · Off-MOO-Bench average | average rank: 5.02 ± 0.03 -> 8.77 ± 0.21 | Hurt |
| Add GradNorm to the end-to-end model (ablation) | End-to-End · Off-MOO-Bench average | average rank: 7.32 ± 0.01 -> 8.34 ± 0.01 | Hurt |
| Add PcGrad to the end-to-end model (ablation) | End-to-End · Off-MOO-Bench average | average rank: 7.32 ± 0.01 -> 7.51 ± 0.14 | Hurt |
| Add GradNorm to the multi-head model (ablation) | Multi-Head · Off-MOO-Bench average | average rank: 7.00 ± 0.38 -> 9.63 ± 0.04 | Hurt |
| Add PcGrad to the multi-head model (ablation) | Multi-Head · Off-MOO-Bench average | average rank: 7.00 ± 0.38 -> 8.09 ± 0.20 | Hurt |
| Use data pruning to focus training on higher-scoring solutions (ablation) | Multi-Head model · RE21 | hypervolume | Helped |
Surprising findings
- No offline MOO method surpasses the training-set best solution on combinatorial (MOCO) tasks.
- Training on all data often yields inferior performance; data pruning of low-quality solutions was needed to get an accurate surrogate.
Limitations
- No single method stands out; overall gains are modest.
- Constraint handling is simplistic (infeasible solutions discarded).
- Surrogate over-estimation persists even with data pruning.
Experimental setup
Datasets: DTLZ ZDT Omni-test VLMOP NAS-Bench-201 C-10/MOP IN-1K/MOP MuJoCo PG-MORL policies TSP/CVRP/KP instances Markowitz portfolio stock data GSK3beta/JNK3 molecule latent space Regex/ZINC/RFP protein sequences
Benchmarks: Off-MOO-Bench
Metrics: hypervolume average rank
Baselines: End-to-End Multi-Head Multiple Models + COMs Multiple Models + RoMA Multiple Models + IOM Multiple Models + ICT Multiple Models + Tri-Mentoring MOBO (qNEHVI) MOBO-qParEGO MOBO-JES
Builds on: Design-Bench COMs RoMA IOM NSGA-II
Closest papers by meaning
- ParetoFlow: Guided Flows in Multi-Objective Optimization (ICLR 2025 · similarity 0.88)
- Offline Model-Based Optimization by Learning to Rank (ICLR 2025 · similarity 0.87)
- Pareto Front-Diverse Batch Multi-Objective Bayesian Optimization (AAAI 2024 · similarity 0.86)
- Preference-Guided Diffusion for Multi-Objective Offline Optimization (NeurIPS 2025 · similarity 0.86)
- Runtime Analysis of the SMS-EMOA for Many-Objective Optimization (AAAI 2024 · similarity 0.86)
- ROOT: Rethinking Offline Optimization as Distributional Translation via Probabilistic Bridge (NeurIPS 2025 · similarity 0.86)
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.