In my experience coordinating and analyzing unmanned combat aerial vehicle (UCAV) operations, one of the most persistent and complex challenges is the problem of target assignment for a drone formation. When multiple autonomous aircraft must engage multiple targets, the decision-making process transcends simple “who shoots what” logic. We must account for a plethora of factors, many of which are inherently qualitative and uncertain. Traditional operations research methods often stumble when trying to quantify concepts like “threat level,” “mission criticality,” or “platform reliability.” It was this gap between crisp mathematical models and fuzzy operational reality that led me to explore and advocate for the application of fuzzy inferential methods. This approach provides a robust framework to navigate the ambiguities of modern aerial warfare, transforming commander’s intent and experiential knowledge into actionable, optimized assignment plans for a cohesive drone formation.
The core of the problem lies in the multi-criteria nature of the decision. Assigning assets from a drone formation is not merely about maximizing kill probability; it’s a balancing act. Consider a scenario where a drone formation of four heterogeneous UCAVs is tasked with neutralizing four distinct targets. We must evaluate each potential drone-target pair against several often-conflicting qualitative conditions:
- Pre-engagement Threat: How severe is the target’s air defense network to a specific drone type? This isn’t a fixed number but a judgment based on sensor range, weapon lethality, and electronic warfare capability.
- Expected Damage Effect: If engaged, what is the likelihood of achieving the desired level of destruction (e.g., mission kill vs. total destruction)? This depends on warhead type, target hardening, and attack angle.
- Platform Reliability & Survivability: What is the assessed probability of the drone surviving the engagement phase based on its stealth characteristics, maneuverability, and known system statuses?
Conventional methods like Linear Programming or simple weighted sums struggle because they require precise numerical inputs for all criteria. How does one definitively state that Threat is a “7” and Damage Effect is an “8”? Fuzzy inference, rooted in Lotfi Zadeh’s fuzzy set theory, allows us to work with these gradations of truth. Instead of an element strictly belonging or not belonging to a set (e.g., “high threat”), it has a degree of membership, a value between 0 and 1. This mirrors human reasoning, where a target can be “somewhat threatening” or “very valuable.” For a drone formation commander, this translates to a much more natural and expressive language for defining mission constraints and priorities.
The methodology I employ begins by constructing a structured comparison for each qualitative condition. For a given condition \(k\) (e.g., “Pre-engagement Threat”) and a specific target \(t_j\), we compare every pair of drones in the drone formation. This pairwise comparison generates a preference degree. Let \(u_i\) and \(u_l\) be two drones in our formation of \(n\) assets. We define a comparison degree \(e_{il}^j\) for target \(j\) under condition \(k\):
- \(e_{il}^j = 1\) if \(u_i\) is strictly superior to \(u_l\) for attacking target \(j\) considering condition \(k\).
- \(e_{il}^j = 0.5\) if \(u_i\) and \(u_l\) are equally suited.
- \(e_{il}^j = 0\) if \(u_i\) is inferior to \(u_l\).
Performing this for all pairs yields a comparison matrix for target \(j\) and condition \(k\):
$$
E^{jk} = [e_{il}^j]_{n \times n}, \quad i,l \in \{1,2,…,n\}
$$
To aggregate this pairwise data into a score for each drone, we sum the rows of this matrix. For drone \(u_i\) against target \(j\) under condition \(k\):
$$
E_i^{jk} = \sum_{l=1}^{n} e_{il}^j
$$
A higher \(E_i^{jk}\) indicates a more preferable drone for that specific target-condition pair. To normalize these scores into a usable preference metric, we identify the maximum row sum \(E_{max}^{jk}\) and compute a normalized preference value \(q_{ij}^k\). A simple yet effective normalization I frequently use is:
$$
q_{ij}^k = \left\lfloor 10 \times \frac{E_i^{jk}}{E_{max}^{jk}} \right\rfloor / 10
$$
where \(\lfloor \cdot \rfloor\) denotes the floor function. This scales the best performer to 1.0 and others proportionally, resulting in a value between 0 and 1. Repeating this process for all \(n\) drones and \(m\) targets gives us the fundamental building block: the Condition-Specific Preference Matrix \(Q^k\) for each qualitative condition \(k\).
$$
Q^k = [q_{ij}^k]_{n \times m}, \quad k = 1, 2, …, K
$$
The following table illustrates the conceptual output of this step for a single condition (e.g., Threat) across a drone formation of 4 UCAVs and 4 targets.
| Drone \ Target | \(t_1\) | \(t_2\) | \(t_3\) | \(t_4\) |
|---|---|---|---|---|
| \(u_1\) | 0.85 | 0.60 | 1.00 | 0.75 |
| \(u_2\) | 0.70 | 1.00 | 0.80 | 0.90 |
| \(u_3\) | 1.00 | 0.75 | 0.65 | 0.60 |
| \(u_4\) | 0.55 | 0.80 | 0.70 | 1.00 |
Now, the real challenge begins: synthesis. A drone formation commander rarely cares about one condition in isolation. The art of command involves weighing the importance of threat versus damage versus survivability. This is where fuzzy inference truly shines. We introduce two critical vectors:
- Condition Weight Vector (\(\Theta\)): This defines the relative importance of each qualitative condition for the overall mission. For example, in a high-risk penetration mission, survivability (\(\Theta_3\)) might be weighted higher than damage effect (\(\Theta_2\)). These weights are normalized.
$$
\Theta = (\Theta_1, \Theta_2, …, \Theta_K)^T, \quad \sum_{k=1}^{K} \Theta_k = 1, \quad \Theta_k \in [0,1]
$$
- Condition Importance Level Vector (\(\Phi\)): This is a more nuanced concept. It defines the desired “ideal” level for each condition on a scale (e.g., from 0 to 1). For instance, we may ideally want minimal threat (desired level ~0), maximum damage (desired level ~1), and maximum reliability (desired level ~1). This vector \(\Phi = (\Phi_1, \Phi_2, …, \Phi_K)\) anchors our optimization toward a commander-defined ideal point.
The goal is to find, for each drone-target pair \((i,j)\), a vector of relative membership degrees \(\Omega_{ij} = (\omega_{ij}^1, \omega_{ij}^2, …, \omega_{ij}^K)^T\). Each \(\omega_{ij}^k\) represents the degree to which pair \((i,j)\) satisfies the combined importance of condition \(k\) in the context of all conditions. We find this by minimizing the weighted squared distance between the achieved condition values (\(q_{ij}^k\)) and their ideal importance levels (\(\Phi_k\)), weighted by the condition’s importance (\(\Theta_k\)). The objective function for pair \((i,j)\) is:
$$
\min F(\Omega_{ij}) = \min \sum_{k=1}^{K} (\omega_{ij}^k)^2 \left[ \Theta_k (q_{ij}^k – \Phi_k) \right]^2
$$
$$
\text{subject to: } \sum_{k=1}^{K} \omega_{ij}^k = 1, \quad \omega_{ij}^k \in [0,1]
$$
Using Lagrange multiplier techniques to solve this constrained optimization yields an elegant closed-form solution for the relative membership degree of condition \(k\) for drone-target pair \((i,j)\):
$$
\omega_{ij}^k = \frac{ \left[ \Theta_k (q_{ij}^k – \Phi_k) \right]^{-2} }{ \sum_{c=1}^{K} \left[ \Theta_c (q_{ij}^c – \Phi_c) \right]^{-2} }
$$
This formula is powerful. It automatically assigns higher membership \(\omega_{ij}^k\) to conditions where the drone-target performance \(q_{ij}^k\) is close to its ideal \(\Phi_k\), and does so in proportion to the condition’s weight \(\Theta_k\). The final, comprehensive score for each drone-target pair \((i,j)\) is the characteristic value \(G_{ij}\), obtained by a weighted sum of the condition memberships using the condition weights as coefficients:
$$
G_{ij} = \sum_{k=1}^{K} \Theta_k \cdot \omega_{ij}^k
$$
Aggregating all these scores forms the Global Assignment Suitability Matrix \(G\).
$$
G = [G_{ij}]_{n \times m}
$$
This matrix \(G\) encapsulates all fuzzy judgments, condition weights, and ideal points into a single, crisp score for every possible assignment within the drone formation. A higher \(G_{ij}\) indicates a more suitable match according to the defined multi-criteria fuzzy logic. The final step is to select the assignment that maximizes the total suitability of the formation. This becomes a classic assignment problem, solvable via algorithms like the Hungarian method:
$$
\begin{aligned}
& \max \sum_{i=1}^{n} \sum_{j=1}^{m} G_{ij} \cdot x_{ij} \\
& \text{subject to:} \\
& \sum_{i=1}^{n} x_{ij} = 1 \quad \forall j \text{ (Each target assigned to one drone)} \\
& \sum_{j=1}^{m} x_{ij} = 1 \quad \forall i \text{ (Each drone assigned to one target)} \\
& x_{ij} \in \{0, 1\}
\end{aligned}
$$
To solidify understanding, let’s walk through a detailed example consistent with the principles derived from operational analysis. Suppose our drone formation has \(n=4\) UCAVs, and we have \(m=4\) targets. We consider \(K=3\) qualitative conditions: C1 (Threat), C2 (Damage), C3 (Reliability). Assume we have already derived the three condition-specific preference matrices \(Q^1, Q^2, Q^3\) through pairwise comparisons as described earlier.
Let the commander-defined parameters be:
- Condition Weights: \(\Theta = (0.35, 0.40, 0.25)^T\) (Damage is slightly prioritized).
- Ideal Importance Levels: \(\Phi = (0.0, 1.0, 1.0)\) (Minimize threat, maximize damage and reliability).
Now, let’s calculate the suitability \(G_{11}\) for Drone 1 to Target 1. Assume the extracted preference scores from our matrices are:
\(q_{11}^1 = 0.85\) (Threat score),
\(q_{11}^2 = 0.70\) (Damage score),
\(q_{11}^3 = 0.90\) (Reliability score).
First, we compute the squared weighted distances for each condition \(k\):
$$
\begin{aligned}
d_1 &= [0.35 \times (0.85 – 0.0)]^2 = (0.2975)^2 = 0.0885 \\
d_2 &= [0.40 \times (0.70 – 1.0)]^2 = (0.40 \times -0.30)^2 = (-0.12)^2 = 0.0144 \\
d_3 &= [0.25 \times (0.90 – 1.0)]^2 = (0.25 \times -0.10)^2 = (-0.025)^2 = 0.000625
\end{aligned}
$$
Then, compute the relative membership degrees \(\omega_{11}^k\):
$$
\begin{aligned}
\omega_{11}^1 &= \frac{1/0.0885}{1/0.0885 + 1/0.0144 + 1/0.000625} = \frac{11.299}{11.299 + 69.444 + 1600} \approx 0.0067 \\
\omega_{11}^2 &= \frac{1/0.0144}{1680.743} \approx \frac{69.444}{1680.743} \approx 0.0413 \\
\omega_{11}^3 &= \frac{1/0.000625}{1680.743} \approx \frac{1600}{1680.743} \approx 0.9520
\end{aligned}
$$
Finally, the characteristic value \(G_{11}\) is:
$$
G_{11} = (0.35 \times 0.0067) + (0.40 \times 0.0413) + (0.25 \times 0.9520) \approx 0.0023 + 0.0165 + 0.2380 \approx 0.2568
$$
Repeating this process for all 16 drone-target pairs yields the complete matrix \(G\). A hypothetical resulting matrix might look like this:
| Drone \ Target | \(t_1\) | \(t_2\) | \(t_3\) | \(t_4\) |
|---|---|---|---|---|
| \(u_1\) | 0.2568 | 0.4150 | 0.7210 | 0.3321 |
| \(u_2\) | 0.5010 | 0.2987 | 0.4502 | 0.6895 |
| \(u_3\) | 0.7105 | 0.5233 | 0.3819 | 0.2944 |
| \(u_4\) | 0.3344 | 0.6555 | 0.5022 | 0.4100 |
Solving the assignment problem for matrix \(G\) above yields the optimal one-to-one assignment: \(u_1 \rightarrow t_3\), \(u_2 \rightarrow t_4\), \(u_3 \rightarrow t_1\), \(u_4 \rightarrow t_2\). This solution maximizes the total suitability score \(G_{13} + G_{24} + G_{31} + G_{42}\) for the entire drone formation.
The advantages of this fuzzy inferential approach for drone formation management are manifold. Firstly, it seamlessly integrates subjective, expert knowledge (via pairwise comparisons and the setting of \(\Theta\) and \(\Phi\)) into a rigorous mathematical framework. Secondly, it is highly adaptable; adding a new qualitative condition simply involves creating another preference matrix \(Q^{new}\) and adding a weight \(\Theta_{new}\). Thirdly, the model’s transparency is a significant operational benefit. Unlike a “black-box” neural network, the commander can trace the logic: from pairwise comparisons to condition membership, and finally to the global score. This builds trust in the automated decision-support system.
In practice, implementing this for a dynamic drone formation requires an efficient software solver. The pairwise comparison step, while conceptually simple, can be automated using rule-based systems or even lightweight machine learning classifiers that output preference degrees. The subsequent calculations for \(\omega_{ij}^k\) and \(G_{ij}\) are computationally inexpensive. The Hungarian algorithm for the final assignment is well-known for its efficiency (\(O(n^3)\) complexity), making this approach feasible for real-time or near-real-time replanning as the battle space evolves.
Looking forward, the fusion of fuzzy inference with other AI techniques presents exciting avenues. A fuzzy-based pre-screening module could generate high-quality candidate assignments for a more complex, nonlinear optimizer that also incorporates detailed kinematic and fuel constraints. Furthermore, the condition weights \(\Theta\) need not be static; they could be dynamically adjusted by a higher-level meta-controller based on the mission phase (e.g., ingress, attack, egress). This creates a hierarchical, intelligent decision-making architecture for autonomous drone formations.
In conclusion, the application of fuzzy inferential methods resolves a critical impedance mismatch in autonomous systems design: the mismatch between human-like, qualitative operational reasoning and the numerical world of optimization algorithms. By providing a formal bridge, it enables us to craft target assignment strategies for a drone formation that are not only mathematically sound but also aligned with the nuanced, experience-based priorities of military command. This synthesis of human judgment and machine calculation is, in my view, essential for unlocking the full collaborative potential of intelligent drone formations in complex, contested environments.