Robots that Talk Less to Each Other but Learn Better: The Secret Is in Compression

Imagine a thousand robots that need to solve a problem together, but can only communicate with very short messages. How do they coordinate? A new study shows that the answer lies in data compression and a trick called momentum.

Researchers developed a special algorithm for distributed optimization, meaning when many agents (could be drones, sensors, or smart devices) work together without a leader. The problem? Sending messages between devices consumes a lot of energy and slows everything down.

The solution is called compression: instead of sending very long numbers, each robot sends only a simplified version, like summarizing a thousand-word essay into three sentences. But there's a risk: this loses precision and the group could head in the wrong direction.

This is where momentum comes in, a technique already used in machine learning. Think of it like muscle memory: if you've been going right for the last few steps, the right direction is probably still there. The algorithm uses this "memory" to compensate for compression errors.

Tests show the system works: robots reach the best solution even with minimally compressed messages, saving energy and time. A step forward for connected device networks, from smart cities to delivery drones.

The developed algorithm belongs to the class of zero-order methods, which do not require computing the exact gradient of the objective function. This is crucial when the gradient is not available in analytical form, as happens with "black-box" functions or when optimization is based on simulations.

The implemented compression is stochastic: instead of sending exact values, each agent quantizes their updates into a finite set of levels. This operation introduces a quantization error that, if not managed, can compromise convergence.

The momentum used is of the heavy-ball type, which combines the current update with a fraction of the previous gradient. Mathematically, the descent direction becomes a weighted average of past gradients, where the memory weight decreases exponentially.

The theoretical results demonstrate two convergence regimes: with diminishing steps, the stationary point is reached exactly; with fixed steps, sublinear convergence is achieved towards a neighborhood of the stationary point, whose width depends on the compression level. This means a quantifiable trade-off exists between communication and accuracy.

Practical applications include distributed training of neural networks on edge devices, coordinated control of drone swarms, and optimization of wireless sensor networks with limited bandwidth.