⚡ electrical engineering

Data-Driven Distributed Optimization via Aggregative Tracking and Deep-Learning

This paper proposes a novel distributed data-driven optimization algorithm that combines neural network-based learning, optimization routines, and aggregative tracking to solve cooperative minimization problems with unknown objectives, proving its linear convergence to a neighborhood of the optimum under strongly convex conditions.

Original authors: Riccardo Brumali, Guido Carnevale, Giuseppe Notarstefano

Published 2026-01-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Riccardo Brumali, Guido Carnevale, Giuseppe Notarstefano

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

Imagine a group of friends trying to organize a massive potluck dinner. Their goal is to make the meal as delicious as possible for everyone. However, there's a catch: no one knows the exact recipe for the perfect dish, and they can't taste the whole meal at once. They can only take one tiny bite of their own dish at a time to see how it tastes.

This is the real-world problem the paper tackles. The authors, Riccardo Brumali, Guido Carnevale, and Giuseppe Notarstefano, propose a new way for a team of "agents" (like robots, computers, or even people) to work together to find the best solution when they don't have the full rulebook. They call their solution DELTA (DEep-Learning aggregative TrAcking).

Here is how DELTA works, broken down into three simple roles that every team member plays simultaneously:

1. The "Gourmet Chef" (The Learning Part)

Since the friends don't know the perfect recipe, they need to guess. Each friend has a neural network (think of it as a very smart, digital apprentice chef).

The Job: Every time the friend takes a tiny bite (a single data sample), the apprentice chef looks at the taste and updates its internal "recipe book."
The Trick: Usually, guessing a recipe requires tasting the food many times to get it right. But this apprentice is special. It learns to predict the direction in which the taste will improve (the gradient) using just that one single bite. It's like a chef who can tell you exactly how much salt to add just by tasting a single grain of rice.

2. The "Team Captain" (The Optimization Part)

Once the apprentice chef tells the friend, "Hey, if you move your dish slightly to the left, it will taste better," the friend needs to actually move the dish.

The Job: This part takes the advice from the apprentice and adjusts the friend's local decision (like moving a robot or changing a setting) to get closer to the best possible meal.
The Challenge: The friend doesn't know the whole picture. They only know their own dish. To make the entire potluck better, they need to know how their dish affects the group's overall flavor.

3. The "Messenger" (The Tracking Part)

This is the glue that holds the team together. Since no one can see the whole table, they need to figure out what the "average flavor" of the whole group is.

The Job: Each friend whispers to their immediate neighbors, "Here is my dish's flavor." The neighbors pass that message along. Through this chain of whispers, every friend eventually builds a mental map of the entire group's average flavor without ever leaving their spot.
The Result: This allows every friend to know how their specific action impacts the global goal, even though they only talk to their neighbors.

How They Work Together (The Magic of Speed)

The genius of DELTA is that these three things happen at the same time and at different speeds, like a well-rehearsed orchestra:

The Messengers (Tracking) move very fast, quickly figuring out the group average.
The Captains (Optimization) move at a medium pace, adjusting the decisions.
The Chefs (Learning) move a bit slower, carefully refining their understanding of the taste based on the new bites.

Because they move at different speeds, the system doesn't get confused. The fast messengers give the captains the info they need, and the captains give the chefs the new data points they need to learn.

The Result: Getting "Close Enough"

The paper proves mathematically that this system works incredibly well.

The Guarantee: If the problem is "convex" (meaning there is one clear best solution, like a single valley in a mountain range), the team will converge to a spot very close to the perfect solution.
The Limit: How close they get depends on how good the "apprentice chefs" (the neural networks) are at guessing. If the networks are perfect, the team finds the exact solution. If the networks are just "pretty good," the team finds a solution that is almost perfect. The paper shows that the error is directly tied to the accuracy of these neural networks.

Why This Matters (According to the Paper)

The authors tested this with a simulation of 20 agents. They showed that:

DELTA works: The team successfully found a near-perfect solution.
It's efficient: It only needs one "taste test" per step, which is crucial when testing is expensive or slow.
It's robust: If the "recipes" (the environment) suddenly change, the team doesn't crash. The apprentice chefs quickly re-learn the new tastes, and the team adapts to the new best solution without starting over.

In short, DELTA is a clever way for a group of strangers to solve a complex puzzle together, even when they can only see a tiny piece of the puzzle at a time and have to learn the rules as they go.

Technical Summary: Data-Driven Distributed Optimization via Aggregative Tracking and Deep-Learning (DELTA)

Problem Formulation
The paper addresses the challenge of distributed optimization in an aggregative framework where a network of $N$ agents cooperatively minimizes the sum of local cost functions. Unlike standard distributed optimization, each agent's local cost $f_i$ depends not only on its own decision variable $x_i$ but also on a global aggregative variable $\sigma(x)$ , defined as the average of local contributions $\phi_i(x_i)$ .

The specific scenario considered is data-driven and unknown:

The functions $f_i$ are entirely unknown to the agents.
Agents can only obtain a single feedback sample of their local cost per iteration (a "1-point" evaluation) at a chosen point $(u_{x,i}, u_{\sigma,i})$ .
Global quantities (the aggregative variable $\sigma(x)$ and the derivative of the global cost with respect to it) are unavailable locally.

This setup is motivated by applications in cooperative robotics and personalized optimization, where cost functions may represent user dissatisfaction or physical constraints that are difficult to model analytically but can be sampled via sensors or user feedback.

Methodology: The DELTA Algorithm
The authors propose DELTA (DEep-Learning aggregative TrAcking), a distributed algorithm that operates on a single timescale and integrates three concurrent components:

Learning-Oriented Module:
- Each agent $i$ employs a local neural network (NN) $\hat{f}_i(\cdot, \theta_i)$ to approximate the unknown cost function $f_i$ and its gradients.
- Instead of training the network to convergence between optimization steps (which is computationally expensive and requires multiple samples), the NN is updated via a single-step gradient descent using the single cost sample obtained at each iteration.
- To ensure persistent excitation for learning, the algorithm injects dither signals ( $d_{x,i}, d_{\sigma,i}$ ) into the sampling points.
- The NN approximates the descent direction, effectively replacing the need for explicit gradient knowledge.
Optimization-Oriented Routine:
- Agents perform a distributed gradient-like update using the approximated gradients provided by their local NNs.
- The update rule combines the local gradient estimate $\nabla_1 \hat{f}_i$ with a term involving the derivative of the aggregation function $\nabla \phi_i$ and the estimated global gradient component.
Tracking-Oriented Mechanism:
- Since global quantities $\sigma(x)$ and the sum of local gradient derivatives are unavailable, agents maintain auxiliary variables ( $w_i$ and $z_i$ ).
- These variables are updated via a perturbed consensus dynamics (using a doubly stochastic adjacency matrix) to locally reconstruct the missing global aggregative variable and the global gradient component.

Theoretical Analysis and Results
The stability and convergence of DELTA are analyzed using tools from system theory, specifically timescale separation and discrete-time averaging theory. The analysis treats the algorithm as a time-varying two-time-scale system, separating the fast dynamics (tracking and learning) from the slow dynamics (optimization).

Convergence Guarantee: Under assumptions of strong convexity for the global cost and Lipschitz continuity of the functions and their approximations, the authors prove that the algorithm linearly converges to a neighborhood of the optimal solution $x^*$ .
Error Bound: The radius of this neighborhood is explicitly dependent on the parameter $\epsilon$ , which represents the maximum reconstruction error of the neural networks (i.e., the accuracy of the NN approximation at convergence).
Ideal Case: In the theoretical limit where the neural networks achieve perfect estimation ( $\epsilon = 0$ ), the algorithm converges exactly to the optimal solution.
Robustness: The design allows the algorithm to adapt to time-varying cost functions or disturbances without requiring a full restart of the training or tracking processes.

Key Contributions

Novel Integration: The paper introduces a novel scheme that combines neural network learning with distributed aggregative optimization and gradient tracking, specifically tailored for scenarios with only single-point cost evaluations.
Overcoming 1-Point Limitations: Standard zeroth-order 1-point methods often suffer from poor convergence rates or high variance. DELTA addresses this by embedding a dynamic learning strategy (NNs) that learns the descent direction over time without increasing the number of cost evaluations per iteration.
Rigorous Convergence Proof: The authors provide a formal proof of linear convergence in a neighborhood of the optimum, explicitly characterizing the trade-off between the neural network's approximation accuracy and the final optimization error.
Handling Unknown Environments: The framework is designed for completely unknown cost functions, distinguishing it from prior works where parts of the cost function were known or where multiple samples were required per iteration.

Significance and Claims
The paper claims that DELTA offers a practical solution for personalized optimization and cooperative control in unknown or non-stationary environments (e.g., human-robot collaboration). By leveraging the synergy between learning, optimization, and tracking, the algorithm adapts to changes in cost functions dynamically.

The authors note that while a preliminary version of this work existed for time-varying settings with multiple samples and full NN training, the current contribution is technically more demanding: it operates with single samples per iteration and single-step NN training, requiring a more complex integration of learning dynamics analysis with optimization and consensus dynamics. Numerical simulations on a 20-agent network validate the theoretical results, demonstrating convergence to the optimal neighborhood and robustness to sudden changes in local cost functions.