Offline Multi-Task Multi-Objective Optimization

• Offline MTMOO is defined as the simultaneous optimization of conflicting objectives across multiple tasks using non-interactive batch data.
• Key methodologies include constrained batch decomposition, joint scalarization with transfer, and surrogate modeling to approximate well-distributed Pareto fronts.
• Approaches leverage gradient descent, evolutionary algorithms, and RL-based techniques while addressing challenges such as scalability and evaluation fidelity.

Offline Multi-Task Multi-Objective Optimization (MTMOO) refers to the study and development of batch (offline) algorithms for simultaneously optimizing multiple conflicting objectives across multiple tasks, under a setting where data and evaluations are non-interactive and no further environment access occurs during optimization. It generalizes both traditional multi-objective optimization (MOO) and multitask learning/optimization, aiming to discover well-distributed Pareto sets or Pareto manifolds representing explicit trade-offs among competing objectives/tasks. Offline MTMOO spans both continuous and combinatorial domains, and includes gradient-based, evolutionary, and surrogate-assisted paradigms as well as reinforcement learning generalizations.

1. Mathematical Foundations and Pareto Theory

Multi-Task Multi-Objective Optimization is formalized as the search for optimal trade-offs in a parameterized decision space $\Theta\subseteq\mathbb{R}^n$, given $T$ potentially conflicting tasks, each with a differentiable loss $L_t:\Theta\to\mathbb{R}$. The core optimization problem is

$$\min_{\theta\in\Theta} F(\theta) = \left(L_1(\theta),\,L_2(\theta),\dots,L_T(\theta)\right)$$

No single $\theta$ typically minimizes all $L_t$; instead, interest centers on the Pareto set: those $\theta^\star$ not dominated under the following relation:

• Pareto dominance: $\theta^a\preceq \theta^b$ iff $\forall t: L_t(\theta^a)\leq L_t(\theta^b)$ and $\exists t: L_t(\theta^a)<L_t(\theta^b)$.
• (Strong/weak) Pareto optimality: $T$0 is (strongly) Pareto-optimal if no $T$1 dominates it; weakly if no $T$2 exists with all $T$3 (Lin et al., 2019).

For constrained settings, the Pareto condition extends via Fritz-John points: $T$4 not both zero, such that

$T$5

where $T$6 are the constraints. The locus $T$7 (with $T$8 the Fritz–John matrix) characterizes the Pareto manifold (Gupta et al., 2021).

2. Core Offline MTMOO Methodologies

2.1 Constrained Batch Decomposition

A canonical batch strategy is to decompose the Pareto front into $T$9 subregions by selecting $L_t:\Theta\to\mathbb{R}$0 "preference" vectors $L_t:\Theta\to\mathbb{R}$1 on the positive simplex. For each $L_t:\Theta\to\mathbb{R}$2, define the corresponding region $L_t:\Theta\to\mathbb{R}$3. The $L_t:\Theta\to\mathbb{R}$4th subproblem becomes:

$L_t:\Theta\to\mathbb{R}$5

equivalently, as a set of linear inequalities $L_t:\Theta\to\mathbb{R}$6 for all $L_t:\Theta\to\mathbb{R}$7 (Lin et al., 2019).

These $L_t:\Theta\to\mathbb{R}$8 subproblems are solved in parallel, often via a projected/constrained batch gradient descent using KKT duality to compute the optimal descent direction, collecting the set $L_t:\Theta\to\mathbb{R}$9 as a Pareto front approximation (Lin et al., 2019).

2.2 Joint Multitask Scalarization and Transfer

Alternatively, the offline MTMOO problem can be formulated by sampling $$\min_{\theta\in\Theta} F(\theta) = \left(L_1(\theta),\,L_2(\theta),\dots,L_T(\theta)\right)$$0 weight vectors $$\min_{\theta\in\Theta} F(\theta) = \left(L_1(\theta),\,L_2(\theta),\dots,L_T(\theta)\right)$$1 on the task simplex and solving $$\min_{\theta\in\Theta} F(\theta) = \left(L_1(\theta),\,L_2(\theta),\dots,L_T(\theta)\right)$$2 unconstrained scalarizations:

• Weighted sum: $$\min_{\theta\in\Theta} F(\theta) = \left(L_1(\theta),\,L_2(\theta),\dots,L_T(\theta)\right)$$3
• Smoothed Tchebycheff: $$\min_{\theta\in\Theta} F(\theta) = \left(L_1(\theta),\,L_2(\theta),\dots,L_T(\theta)\right)$$4 with softmax aggregation centered at per-task minima (Bai et al., 2024).

Instead of independent optimization, "multi-task gradient descent" applies transfer between iterates via a matrix $$\min_{\theta\in\Theta} F(\theta) = \left(L_1(\theta),\,L_2(\theta),\dots,L_T(\theta)\right)$$5:

$$\min_{\theta\in\Theta} F(\theta) = \left(L_1(\theta),\,L_2(\theta),\dots,L_T(\theta)\right)$$6

Accelerated convergence is established under strong convexity and smoothness, with spectral convergence factor $$\min_{\theta\in\Theta} F(\theta) = \left(L_1(\theta),\,L_2(\theta),\dots,L_T(\theta)\right)$$7 (single-task) (Bai et al., 2024).

2.3 Surrogate and Meta-learning Approaches

For expensive, complex, or black-box multi-task MO functions, LLM-based surrogates such as Q-MetaSur tokenize the MTMOO instance (metadata plus input vector) and regress the vectorial objectives as sequences. This sequence-to-sequence setup is trained by supervised teacher forcing with priority-weighted cross-entropy (PWCE), followed by offline RL fine-tuning with Q-learning (ILQL) and conservative Q-regularization, utilizing explicit rewards tied to normalized RMSE and bit-level correctness (Zhang et al., 17 Dec 2025).

At inference, surrogate prediction replaces true evaluation within any underlying evolutionary optimizer; advantage-guided decoding increases robustness to out-of-data samples.

2.4 Evolutionary and Multifactorial Optimization

Multifactorial evolutionary algorithms (MFEA) enable offline MTMOO by maintaining a single population, each individual annotated with a skill factor denoting task specialization (Yuan et al., 2017, Guo et al., 2023). Operators include selective mating (crossover when skill factors match or random threshold is met, else mutation) and vertical cultural transmission for efficient "skill" inheritance.

Selection leverages strategy pools (vector-angle, tournament, grid-based) to maintain diversity and convergence across high-dimensional objectives (Guo et al., 2023).

3. Representative Algorithms and Their Properties

Algorithm/Class	Core Technique	Pareto Coverage	Key Features
Pareto MTL (Lin et al., 2019)	Constrained QP	Well-distributed	Batch, subproblem parallelism
MT²O (Bai et al., 2024)	MT Transfer GD	Dense	Fast convergence, scalarization/transfers
Q-MetaSur (Zhang et al., 17 Dec 2025)	LLM surrogate	Nearly exact	Unified seq2seq, RL regularization
MOMFEA-MS (Guo et al., 2023)	Multifactorial EA	Diverse	Skill-factor, multi-selection
SUHNPF (Gupta et al., 2021)	Double-gradient	Dense manifold	Fritz–John, classifier induction
Policy-regularized MORL (Lin et al., 2024)	RL (actor-critic)	Dense conditional	Pref-conditioned, BC filtering

Each method offers specific advantages: Pareto MTL and MT²O efficiently span Pareto fronts in neural multitask learning, Q-MetaSur enhances data-driven search under expensive black-box evaluations, and MOMFEA-MS achieves robust solutions in high-dimensional, multi-task edge computing scenarios. SUHNPF enables dense Pareto manifold extraction even in the presence of explicit constraints.

4. Experimental Protocols and Benchmarks

Comprehensive benchmarking has utilized both synthetic two-objective landscapes (ZDT1, ZDT2, concave fronts) and realistic MTMOO scenarios:

• MultiMNIST/MultiFashionMNIST (conflicting classification)
• NYUv2 (scene understanding: segmentation, depth, normals)
• CelebA (multi-label, $$\min_{\theta\in\Theta} F(\theta) = \left(L_1(\theta),\,L_2(\theta),\dots,L_T(\theta)\right)$$8)
• Edge computing deployment and offloading (4 objectives per task) (Guo et al., 2023)

Key metrics for Pareto set quality include:

• Hypervolume (HV): total dominated volume; higher HV indicates better approximate Pareto front (Bai et al., 2024).
• Inverted Generational Distance (IGD): mean minimum distance from reference Pareto front (Yuan et al., 2017, Zhang et al., 17 Dec 2025).
• Mean Standard Score (MSS): task-averaged normalized IGD (Yuan et al., 2017).
• Sparsity (Sp): point-density along front (lower is better).
• Task-specific utility metrics: accuracy, error, RLP, mIoU, etc.

A unifying outcome is that joint or surrogate-driven MTMOO approaches (MT²O, Q-MetaSur, MOMFEA-MS) outperform single-task or naive scalarization baselines in both convergence and coverage, especially in high-similarity or partially overlapping multitask settings (Yuan et al., 2017, Bai et al., 2024, Guo et al., 2023).

5. Specializations: Offline Batch, RL, and High-dimensional MTMOO

Offline MTMOO methods operate entirely over pre-collected datasets (or batch-evaluated surrogates), with no environment access during optimization. In offline RL, policy-regularized multi-objective actor-critic setups embed user preferences as inputs, solve scalarized Bellman equations with a regularization term ensuring proximity to observed behavior, and filter "preference-inconsistent" trajectories via cosine alignment of empirical returns (Lin et al., 2024). RL-specific challenges include trade-off-dependent behavior cloning weights (tuned adaptively by introducing them as preference dimensions), and conditional value function estimation.

For high-dimensional and multi-user resource allocation (e.g., edge computing), MOMFEA-MS treats deployment and offloading as coupled MTMOO tasks, addresses the four-objective regime using grid/tournament/angle selection pooling to retain solution diversity, and quantifies performance across all combinations (Guo et al., 2023).

6. Theoretical Analysis and Limitations

Several frameworks provide theoretical guarantees. The MT²O iteration contracts at least as fast as single-task descent under standard convexity assumptions (Bai et al., 2024). SUHNPF leverages Fritz–John theory and double-gradient refinement, converging rapidly with only a few thousand determinant evaluations even in 30D settings (Gupta et al., 2021). Essential limitations include the requirements for differentiable objectives/constraints (for gradient-based approaches), increased memory/compute scaling with number of tasks or objectives, and the need for large, representative offline datasets if high-fidelity surrogates are employed (Zhang et al., 17 Dec 2025).

Offline MTMOO is inherently a batch setting; adaptations to online scenarios, non-differentiable/non-convex or discrete variable settings, and tasks with substantial inter-task heterogeneity remain open directions.

7. Outlook, Empirical Evidence, and Emergent Best Practices

Empirical studies uniformly demonstrate that:

• Well-designed MTMOO optimizers achieve denser, better-spread Pareto coverage than naive baselines (Lin et al., 2019, Bai et al., 2024, Zhang et al., 17 Dec 2025, Guo et al., 2023).
• Surrogate modeling (LLM-based or otherwise) enables efficient optimization under tight function evaluation budgets, with meta-learning surrogates (Q-MetaSur) offering strong zero-shot and few-shot task generalization (Zhang et al., 17 Dec 2025).
• Joint or transfer-based optimizers accelerate convergence, especially with structural similarity among tasks (Bai et al., 2024, Yuan et al., 2017).
• Practical instantiations (e.g., edge computing deployment) show that multifactorial evolutionary approaches outperform single-task or task-decoupled alternatives in both convergence and diversity (Guo et al., 2023).

Best practices include:

• Sampling dense, uniform reference weight vectors or preference directions for thorough Pareto set approximation.
• Employing multiple diversity-preserving selection/transfer/operator pools in evolutionary or population-based approaches.
• Utilizing advanced regularization (in RL) or conservative policy/value estimation in offline settings with significant demonstration bias (Lin et al., 2024).
• Adopting token-level representation and RL-style surrogate training for high-dimensional, multi-objective function approximation (Zhang et al., 17 Dec 2025).

Offline MTMOO remains a focal research area for resource allocation, neural multitask modeling, recommendation, and automated system design, with ongoing advances in theoretical, algorithmic, and surrogate modeling components.