arXiv preview of Scaling and Trade-offs in Multi-agent Autonomous Systems

Read the full paper preview →

Designing a large autonomous drone swarm forces the engineer to choose across a very wide design space simultaneously: the platform (speed, sensing range, weapon range, mass), the algorithmic layer (cooperative control, path planning, target allocation), and the numerical-strength parameters (how many agents, how fast they move relative to one another, how quickly they are lost). Because these choices interact in nonlinear ways, there is no closed-form way to predict performance, and a brute-force search over all combinations is intractable.

Our Approach

My collaborators and I performed large-scale agent-based simulations in which each drone is an autonomous agent operating under the chosen platform and algorithm. We ran these simulations across three canonical operational scenarios:

• Swarm-on-swarm battle — two opposing swarms engaging one another,
• Cooperative area search with attrition — a single swarm searching a region while agents are progressively lost, and
• Pursuit of scattering targets — a swarm chasing targets that move to evade it.

For each scenario we swept the key design parameters over many orders of magnitude.

Key Findings

Simple but counterintuitive scaling laws

By applying dimensional analysis and data-scaling, we collapsed the high-dimensional performance data onto low-dimensional scaling functions. These functions are mathematically compact, yet their shape is counterintuitive and, as far as we can tell, cannot be predicted a priori from first principles.

Success–failure boundaries and “effective swarm size”

The scaling laws expose sharp success–failure boundaries in parameter space. At certain break points, performance changes abruptly. We showed that these break points can be re-expressed as a single quantity we call the effective swarm size — a number that folds together agent count, platform parameters, and scenario geometry into one governing variable.

Trade-off quantification

Because the scaling functions are explicit, they let us quantify trade-offs directly: for a fixed mission budget, how does swapping agent count for greater velocity, sensing range, weapon range, or a lower attrition rate shift the probability of mission success? These trade-offs can be read off a single curve rather than requiring a new simulation.

Path-planning loop

We further demonstrated that embedding an optimal path-planning loop inside the swarm algorithm qualitatively improves the governing scaling laws, shifting the effective operating regime to more favourable regions of parameter space.

What We Propose

The methods we developed are scenario-agnostic and highly flexible. Taken together, they provide a route to rapid, budget-aware sizing and algorithm selection for large autonomous swarms: instead of simulating thousands of design variants, the designer can consult a small set of scaling functions that already encode the dominant physics of the problem and flag where the most critical performance transitions lie.