Learning to Optimize via Direct Objective Minimization

22 concepts19 papers2 proposals

When faced with complex, NP-hard, or highly non-convex optimization problems—such as allocating power across a cellular network or configuring thousands of antennas in real time—traditional mathematical solvers hit a wall. Algorithms like the Weighted Minimum Mean Squared Error (WMMSE) or interior-point methods are iterative, computationally expensive, and often too slow for real-time deployment.

To accelerate these decisions, researchers initially turned to supervised deep learning. However, supervised learning introduces a severe paradox: to train a neural network to solve an optimization problem quickly, you must first generate a massive dataset of optimal solutions using the very same slow, traditional solvers you are trying to replace. For large-scale or dynamic systems, generating these ground-truth labels is computationally prohibitive or outright impossible.

A powerful paradigm shifts this approach: direct objective minimization. Instead of training a neural network to mimic a traditional solver, the mathematical optimization objective itself is transformed into a differentiable loss function. By treating the physical or structural objective of the system as the loss, the network learns to discover optimal strategies directly through gradient descent, bypassing the need for labeled data entirely.


The Core Mechanism: The Objective as the Loss

In a traditional supervised setup, a neural network is trained to minimize the mean squared error (MSE) between its predictions and a set of target labels. In direct objective minimization, the target labels disappear. Instead, the network takes the environmental state (such as raw channel measurements or graph topologies) as input, outputs a candidate solution (such as beamforming vectors or power levels), and evaluates that solution by passing it directly into the system's analytical objective function.

[Input: System State] ──> [Neural Network] ──> [Candidate Solution]
                                                       │
                                                       ▼
[Gradient Backpropagation] <── [Loss = -Objective(Solution, State)]

Consider the problem of passive beamforming in Reconfigurable Intelligent Surface (RIS) networks. Caijun Zhong et al. demonstrate this mechanism in Unsupervised Learning for Passive Beamforming by using the negative of the receive signal-to-noise ratio (SNR) as the loss function. Because the formula for SNR is differentiable with respect to the RIS phase shifts, gradients can flow backward through the physical equation to update the network's weights. The network essentially learns to shape the wireless environment by maximizing the physical performance metric of the communication link.

This strategy scales seamlessly to multi-user scenarios where the objective is to maximize total data throughput. For instance, Tao Jiang et al. in Learning to Beamform for Intelligent Reflecting Surface with Implicit Channel Estimate utilize a differentiable sum-rate formula as their loss function. By optimizing directly against the final performance metric, the model learns to map raw, imperfect pilot signals directly to optimal beamforming configurations, completely bypassing the traditional, overhead-heavy bottleneck of explicit channel estimation.


Why It Works: End-to-End Alignment

Direct objective minimization works because it aligns the training process perfectly with the ultimate engineering goal. Supervised learning often suffers from metric mismatch; minimizing MSE on intermediate parameters does not guarantee that the resulting system achieves the highest throughput or lowest latency.

This advantage becomes clear when dealing with hardware constraints or multi-task objectives:


Injecting Structure: Unfolding and Graph Neural Networks

While minimizing an objective function directly eliminates the need for labels, optimizing over highly non-convex landscapes can still cause standard feedforward networks to get trapped in poor local minima. To solve this, researchers pair direct objective minimization with domain-specific network architectures that possess strong inductive biases.

Deep Algorithm Unfolding

Deep unfolding maps the iterations of a classical optimization algorithm onto the layers of a neural network. Instead of learning an arbitrary mapping from scratch, the network retains the structural skeleton of the mathematical algorithm, while its hyperparameters (like step sizes or regularization weights) are optimized end-to-end.

For example, Pierre Ablin et al. demonstrate in Learning step sizes for unfolded sparse coding. that an unfolded sparse coding network can be trained in an unsupervised manner by directly minimizing the original lasso objective function to learn optimal, layer-dependent step sizes. In the wireless domain, Nir Shlezinger et al. in Joint Communications and Sensing Hybrid Beamforming Design via Deep Unfolding unfold iterative beamforming algorithms, using the mathematical objective itself as the loss function to achieve massive inference speedups while maintaining the rigorous performance guarantees of classical methods.

Graph Neural Networks (GNNs)

Wireless networks are inherently topological; links interfere with one another based on their spatial arrangement, and rearranging the indexing of users should not change the optimal resource allocation (a property known as permutation equivariance).

Yifei Shen et al. exploit this structural symmetry in Graph Neural Networks for Scalable Radio Resource Management: Architecture Design and Theoretical Analysis. By training a permutation-equivariant GNN with the negative network sum-rate as the loss function, the authors prove a mathematical equivalence between GNNs and a family of distributed optimization algorithms, allowing the model to generalize to network sizes far larger than those seen during training.

To push this further, Arindam Chowdhury et al. combine both paradigms in Efficient power allocation using graph neural networks and deep algorithm unfolding and ML-aided power allocation for Tactical MIMO. By embedding GNNs directly into the iterations of an unfolded WMMSE algorithm and training via direct sum-rate maximization, the authors demonstrate that the model can actually discover optimization strategies that exceed the performance of the original iterative algorithm, while executing orders of magnitude faster.


Generalization Beyond Wireless Systems

While wireless resource management serves as a primary driver for this paradigm, the concept of training networks by directly minimizing a problem's mathematical objective generalizes across domains.

A clear example of this is found in optimal transport theory. The authors of Meta Optimal Transport apply this self-supervised strategy to predict optimal transport plans directly from input data distributions. By defining the loss function as the optimal transport cost itself, the neural network learns to output highly accurate transport plans that can serve as a "warm-start" for traditional solvers, yielding massive computational speedups in applications where similar optimization problems must be solved repeatedly.

Ultimately, direct objective minimization shifts the role of deep learning from a tool that mimics human-engineered algorithms to an autonomous discovery engine. By grounding the loss function in the underlying physics and mathematics of the problem, neural networks learn to optimize complex systems efficiently, flexibly, and entirely without labels.

Key Foundational Work

Unsupervised Learning for Passive Beamforming (2020) — contributed 2 concept elements to this direction.
Unsupervised Learning for Cellular Power Control (2020) — contributed 1 concept elements to this direction.
Graph Neural Networks for Scalable Radio Resource Management: Architecture Design and Theoretical Analysis (2020) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_unsupervised_objective_based_training_bb533722,
  title = {unsupervised objective-based training},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unsupervised-objective-based-training/rdr_bb533722899a476f94aa2e8f93233a27},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_unsupervised_objective_based_training_bb533722_v1,
  title = {unsupervised objective-based training},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unsupervised-objective-based-training/rdr_bb533722899a476f94aa2e8f93233a27/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake