🤖 machine learning

Amortized Multi-Objective Optimization Across Tasks with Generative Solution Modeling

This paper proposes a novel parametric multi-objective Bayesian optimizer that amortizes optimization costs across continuous task parameters by learning an inverse model through alternating generative solution sampling and acquisition-driven search, enabling direct prediction of Pareto optimal solutions for unseen problems without expensive re-evaluations.

Original authors: Tingyang Wei, Jiao Liu, Abhishek Gupta, Chin Chun Ooi, Puay Siew Tan, Yew-Soon Ong

Published 2026-05-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tingyang Wei, Jiao Liu, Abhishek Gupta, Chin Chun Ooi, Puay Siew Tan, Yew-Soon Ong

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Endless Menu" Dilemma

Imagine you are a master chef trying to create the perfect dish. But there's a catch: you have to cook for thousands of different customers, each with slightly different tastes (some like it spicier, some sweeter, some with more salt).

In the real world, this is like Expensive Multi-Objective Optimization (EMOP).

The Chef: An algorithm trying to find the best solution.
The Dishes: Solutions to a problem (like designing a robot arm or a solar panel).
The Taste Test: Evaluating the solution. In this paper, these "taste tests" are expensive. They might take hours of computer time, cost a lot of money, or require physical experiments.

Usually, if a customer changes their mind slightly (a new "task parameter"), the chef has to start cooking from scratch, tasting every dish again to find the new perfect balance. If there are infinite variations of customers, this is impossible. You'd run out of time and money before you could serve everyone.

The Paper's Solution: The "Smart Recipe Book"

The authors propose a new way to cook. Instead of tasting every single dish for every single customer, they build a Smart Recipe Book (an "Inverse Generative Model").

Once they have tasted enough dishes for a few representative customers, they learn a pattern. They realize: "Ah, when the customer wants it spicier, we just need to add a pinch more chili and reduce the sugar by a tiny bit."

With this book, if a brand new customer walks in with a unique taste preference the chef has never seen before, the chef doesn't need to cook or taste anything new. They just look at the book, calculate the recipe, and serve the perfect dish immediately.

How They Built the Book: The "Two-Step Dance"

The paper describes a method called PMT-MOBO that builds this book using a clever "two-step dance" that alternates back and forth:

Step 1: The "Scout" (Acquisition-Driven Search)

Think of this as a scout exploring a vast forest to find the best trees.

The scout uses a map (called a Task-Aware Gaussian Process) that knows the forest isn't just one big empty space; it's a family of connected forests.
Because the scout knows that "Forest A" and "Forest B" are related, finding a great tree in Forest A helps the scout guess where good trees might be in Forest B.
This step is efficient but can sometimes get stuck in a local area, missing the truly amazing trees hidden elsewhere.

Step 2: The "Dreamer" (Generative Solution Sampling)

Think of this as a dreamer who has seen enough great trees to imagine what a perfect tree should look like.

The dreamer uses a Generative Model (like a VAE or a Diffusion Model—think of them as advanced AI artists).
Based on the "elite" trees the Scout found, the Dreamer generates a whole batch of new candidate trees that look promising.
This helps the team jump out of local areas and explore parts of the forest the Scout might have missed.

The Magic: The paper alternates between the Scout (who finds real data) and the Dreamer (who imagines new possibilities). The Scout's data trains the Dreamer, and the Dreamer's ideas give the Scout new places to look. This creates a feedback loop that learns the "shape" of the perfect solutions very quickly.

Why It's Better Than Old Methods

It Learns Relationships: Old methods treated every customer as a stranger. This method knows that customers are related. It uses math (specifically Task-Aware Gaussian Processes) to prove that by sharing information between related tasks, you learn faster. It's like realizing that if you know how to drive a sedan, learning to drive a similar SUV is much easier than learning to drive a truck from scratch.
It Handles "Infinite" Customers: Because the system learns a general rule (the inverse model), it can instantly predict the best solution for a customer with a preference it has never seen before. No expensive re-testing is needed.
It's Robust: The paper shows that if you only use the Scout, you might miss the best spots. If you only use the Dreamer, you might hallucinate bad solutions. But doing both together gives the best results.

The Results: What They Found

The authors tested this on:

Synthetic puzzles: Mathematical problems designed to be tricky.
Real-world scenarios:
- Designing Lamps (balancing stability, cost, and light).
- Designing Solar Rooftops (balancing morning vs. evening energy capture).
- UAV Drones (balancing navigation accuracy vs. battery life).
- Magnetic Sifters (separating cells in medicine).

The Outcome: Their "Smart Recipe Book" (PMT-MOBO) consistently found better solutions faster than existing methods. Most importantly, when they tested it on completely new scenarios (unseen customers), the book worked perfectly, predicting the best designs without needing to run a single new expensive experiment.

In a Nutshell

This paper introduces a system that stops solving expensive problems one by one. Instead, it learns the "grammar" of the problem across many variations. Once it understands the grammar, it can instantly write the perfect solution for any new variation, saving massive amounts of time and money. It does this by constantly switching between carefully exploring the real world and creatively imagining new possibilities.

Technical Summary: Amortized Multi-Objective Optimization Across Tasks with Generative Solution Modeling

1. Problem Formulation

Real-world applications often require solving families of Expensive Multi-Objective Optimization Problems (EMOPs) under varying operational conditions. These are formulated as Parametric Expensive Multi-Objective Optimization Problems (P-EMOPs), where a continuous task parameter $\theta \in \Theta$ defines a distinct optimization instance $F(x, \theta)$ .

The core challenge addressed is that the continuous task parameter space contains infinitely many distinct problems. Traditional Multi-Objective Bayesian Optimization (MOBO) methods, which search for finite sets of Pareto optimal solutions for each task individually, become computationally infeasible when applied repeatedly across this infinite space due to the high cost of objective function evaluations. The goal is to amortize the optimization effort across the continuous task-preference space, enabling the direct prediction of optimized solutions for unseen tasks without re-evaluation.

2. Methodology

The paper proposes a Parametric Multi-Task Multi-Objective Bayesian Optimizer (PMT-MOBO) that learns an inverse generative model $M(\theta, \lambda)$ mapping task parameters $\theta$ and preference vectors $\lambda$ directly to optimized solutions. The framework operates via an alternating optimization loop comprising two complementary phases:

A. Task-Aware Surrogate Modeling (PMT-MOBO)

The method extends standard MOBO by incorporating task-aware Gaussian Processes (GPs). Unlike single-task GPs, these models take the joint input of the decision variable $x$ and the task parameter $\theta$ .

Mechanism: For a given task instance $\theta_k$ , the $m$ -th objective is modeled as $f_m(\cdot, \theta_k) \sim \mathcal{GP}_m(\tilde{\mu}_m, \tilde{\kappa}_m)$ .
Acquisition: The acquisition function is scalarized using a preference vector $\lambda$ and optimized to select the next evaluation point. By conditioning on $\theta$ , the model explicitly leverages inter-task correlations (synergies) to guide the search.

B. Conditional Generative Solution Sampling

To complement the acquisition-driven search and prevent the sampler from getting trapped in low-quality regions, the framework employs a conditional generative model (instantiated as either a Variational Autoencoder (VAE) or a Denoising Diffusion Probabilistic Model (DDPM)).

Training: The model learns the distribution of "elite" solutions conditioned on the task parameter $\theta$ and preference $\lambda$ , denoted as $p_\phi(x | \lambda, \theta)$ .
Sampling: In the generative phase, a pool of candidate solutions is sampled from this distribution. These candidates are then scored using the task-aware acquisition function, and the best candidate is selected for evaluation.

C. Alternating Optimization Loop

The framework alternates between Acquisition-Driven Search (optimizing the task-aware acquisition function directly) and Generative Solution Sampling (sampling from the learned elite distribution and selecting the best). This loop updates the task-aware GPs after every evaluation and retrains the generative model periodically using the collected elite data.

3. Theoretical Contributions

The paper provides a theoretical justification for the accelerated convergence of the proposed method compared to single-task baselines:

Regret Bounds: Using the Lower Confidence Bound (LCB) acquisition function, the authors derive Bayes Regret bounds.
Maximum Information Gain (MIG): Theorem 2 demonstrates that the task-aware kernel in PMT-MOBO reduces the posterior covariance for a target task by leveraging observations from related tasks. This reduction leads to a strictly smaller Maximum Information Gain ( $\tilde{\gamma}_{T,k} \leq \gamma_{T,k}$ ) compared to single-task MOBO (ST-MOBO).
Convergence: Since the regret bound is monotonically increasing with MIG, the reduced MIG implies a tighter upper bound on the Bayes Regret, theoretically proving faster convergence for PMT-MOBO.

4. Experimental Results

The method was evaluated on synthetic benchmarks (DTLZ-1, DTLZ-2, DTLZ-3) and real-world problems (Lamp Generation, Solar Rooftop Generation, Magnetic Sifter Design, UAV Controller Design).

Optimization Performance: The proposed variants, PMT-MOBO-VAE and PMT-MOBO-DDPM, consistently outperformed state-of-the-art single-task MOBO methods (ST-MOBO, qPAREGO, qEHVI) and multi-task baselines (OC-MOBO, PPSL-MOBO) across most benchmarks. Specifically, PMT-MOBO-DDPM achieved the best results on 6 out of 7 benchmarks.
Inverse Model Generalization: The trained inverse models were tested on unseen EMOP tasks (tasks with parameters not seen during training). The proposed models demonstrated superior generalization, achieving higher Hypervolume indicators compared to baselines like PSL-MOBO and CDM-PSL.
Ablation Studies: Experiments confirmed that neither pure acquisition search nor pure generative sampling is sufficient in isolation. The alternating strategy effectively balances diversity (via generative sampling) and precision (via acquisition search).
Sensitivity: The elite ratio parameter $Q$ (percentage of top solutions used for training) was found to be critical; $Q=10\%$ provided the optimal balance between selective pressure and data availability.

5. Significance and Claims

The paper claims to introduce a novel paradigm for solving families of P-EMOPs by shifting from isolated task optimization to learning generalized solution mappings.

Amortization: The primary contribution is the ability to "amortize" the computational cost of repeated multi-objective optimizations into a single inverse generative model pass. This allows for zero-shot prediction of solutions for new operational conditions without expensive re-evaluations.
Inter-Task Synergy: The work demonstrates that explicitly modeling inter-task relationships via task-aware GPs and conditional generative models yields significantly better convergence and generalization than treating tasks in isolation or using standard multi-task approaches that do not model the continuous task-preference manifold.
Distributional Learning: Unlike deterministic hypernetwork approaches (e.g., PPSL-MOBO) that assume a one-to-one mapping, the proposed generative approach models the solution space as a distribution. This allows the framework to capture complex, multi-modal Pareto Fronts and handle the inherent uncertainty in mapping preferences to solutions more robustly.

The authors conclude that while the theoretical regret bounds are established for the task-aware GP component, the integration of generative sampling acts as a heuristic to further enhance search efficiency and solution quality, a claim supported by empirical evidence across diverse engineering and synthetic domains.