The rapid evolution of unmanned aerial vehicles (UAVs) has shifted the design of autonomous systems from single-platform operation toward swarm-level intelligence. In a UAV swarm, each platform is an autonomous decision unit, and the emergence of cooperative behaviors depends on local interactions, self-organization, and adaptive movement. In this article, I investigate the migration-rate dynamics of UAVs from the perspective of evolutionary game theory. I construct a continuous-space spatial prisoner’s dilemma model in which individuals adjust their migration speed and direction based on external feedback and internal learning. I show that the coupling between rate regulation and direction optimization is decisive for the survival and spread of cooperation. My results reveal how different feedback mechanisms at the collective and individual levels shape the spatiotemporal organization of UAV swarms and provide a generalizable framework for distributed cooperative control.
1. Introduction and Motivation
Unmanned aerial vehicles are increasingly used in reconnaissance, target strike, communication relay, and disaster response. In those missions, a single platform often suffers from limited sensing range, energy constraints, and weak anti-jamming ability. A swarm of UAVs can overcome these limitations through information sharing, functional complementarity, and dynamic coordination. However, the coordination capacity of a swarm does not arise from a central command; it emerges from repeated local interactions among individuals. Therefore, understanding the bottom-up mechanisms that drive cooperation in UAV swarms is a central scientific and engineering problem.
Migration is one of the most important behavioral dimensions in a spatial multi-agent system. In a UAV swarm, migration is not simply a change of position. It is a complex decision that combines speed adjustment and direction selection. Individuals need to decide when to decelerate to preserve stable local structures and when to accelerate to explore better neighborhoods. They also need to decide whether to follow the average direction of their neighbors or maintain their own heading. Such decisions are deeply linked to the payoffs that individuals receive from their strategic interactions.
Traditional evolutionary game models on fixed networks often treat migration speed as a constant. That simplification is inconsistent with the realities of UAV swarms. In dynamic mission environments, the speed of a UAV is influenced by task urgency, local density, neighbor alignment, and energy levels. Therefore, I argue that migration rate should be treated as a dynamic variable that responds to both external spatial conditions and internal payoff evaluations.
The central question addressed in this article is: how do different migration-rate mechanisms affect the evolution of cooperation in a UAV swarm? I answer this question by constructing three progressive models. The first model uses directional consistency as an external feedback signal to adjust speed. The second model introduces a Bush–Mosteller reinforcement-learning rule in which speed updates are driven by payoff satisfaction. The third model combines directional memory with local and global order parameters, forming a coupled feedback mechanism. Through these models, I systematically identify the conditions under which adaptive speed control helps drones maintain dense cooperative clusters and resist exploitation by defectors.
2. Theoretical Foundations
2.1 Evolutionary Game Theory and UAV Swarms
Evolutionary game theory provides a natural framework for studying strategic interactions among agents with bounded rationality. A UAV swarm can be viewed as a structured population in which individuals repeatedly play games with their spatial neighbors. The interaction structure can be represented by a continuous-space network, where the neighborhood of each drone is determined by Euclidean distance. In my model, if two drones are within a vision radius \(R_0\), they interact. Since drones move, these neighborhood relations are updated every time step.
The strategic choice of a drone is abstracted into two possible actions: cooperation and defection. Cooperation corresponds to sharing resources, information, or computation with other drones. Defection corresponds to exploiting the swarm while avoiding the cost of cooperation. This dilemma maps directly to the prisoner’s dilemma game. In the weak form of this game, the payoff matrix is:
| Column player C | Column player D | |
|---|---|---|
| Row player C | \(1\) | \(0\) |
| Row player D | \(b\) | \(0\) |
Here \(b > 1\) is the temptation to defect. When two cooperators meet, both receive \(1\). When a cooperator meets a defector, the cooperator receives \(0\) and the defector receives \(b\). When two defectors meet, both receive \(0\). At the individual level, defection is always at least as profitable as cooperation. However, mutual defection leaves everyone worse off than mutual cooperation.
The mapping between UAV swarm concepts and evolutionary game concepts is summarized below:
| UAV Swarm | Evolutionary Game |
|---|---|
| UAV swarm | Structured population |
| Single UAV | Individual player |
| Resource sharing behavior | Cooperation |
| Selfish resource occupation | Defection |
| Spatial movement of UAVs | Migration mechanism |
| Flight speed of a UAV | Migration rate |
| Extra benefit from free-riding | Temptation to defect |
| Task payoff obtained by a UAV | Game payoff |
| Strategy adaptation over missions | Evolutionary dynamics |
| Stable fraction of cooperators | Cooperation level |
2.2 Continuous-Space Network and Neighborhood
Most complex-network models assume fixed links between nodes. This assumption is not appropriate for UAV swarms because the communication and interaction links change as drones move. I therefore adopt a two-dimensional continuous-space model. The positions of drones are described by
\[
p_i(t) = (x_i(t), y_i(t)).
\]
At time \(t\), drone \(j\) is a neighbor of drone \(i\) if
\[
d_{ij}(t) = \sqrt{(x_i(t)-x_j(t))^2 + (y_i(t)-y_j(t))^2} \leq R_0.
\]
The boundary condition is periodic, meaning that a drone leaving one side of the domain re-enters on the opposite side. This avoids artificial edge effects. In all simulations, I fix the domain size \(L=10\), the population size \(N=1000\), and the vision radius \(R_0=0.5\) unless otherwise stated.
2.3 Payoff and Strategy Update
Each drone interacts with all neighbors within \(R_0\). The payoff gained by drone \(i\) from its interaction with neighbor \(j\) is
\[
P_{ij}(t) = s_i(t) s_j(t) + (1-s_i(t)) s_j(t) b,
\]
where \(s_i=1\) for cooperation and \(s_i=0\) for defection. To avoid the influence of heterogeneous degrees, I use normalized payoff:
\[
U_i(t) = \frac{1}{|\Omega_i(t)|} \sum_{j \in \Omega_i(t)} P_{ij}(t),
\]
where \(\Omega_i(t)\) denotes the set of neighbors of drone \(i\) at time \(t\). For strategy update, I use the “rich-follow” rule. Drone \(i\) first identifies the neighbor with the highest payoff:
\[
j^* = \arg\max_{j \in \Omega_i(t)} U_j(t).
\]
If \(U_{j^*}(t) > U_i(t)\), then drone \(i\) adopts the strategy of drone \(j^*\). Otherwise, it keeps its own strategy.
2.4 Migration and Position Update
The position of each drone is updated according to its migration speed \(V_i\) and direction \(\theta_i\):
\[
x_i(t+1) = x_i(t) + V_i(t) \cos\theta_i(t),
\]
\[
y_i(t+1) = y_i(t) + V_i(t) \sin\theta_i(t).
\]
The key contribution of my models is that \(V_i\) and \(\theta_i\) are not externally fixed parameters. They are influenced by local interactions, payoff satisfaction, or both. In the next sections, I present three different mechanisms that determine how these qualitative variables adapt over time.
3. Speed Feedback Based on Neighbor Directional Consistency
3.1 Definition of the Model
In the first model, I assume that each drone adjusts its migration speed according to the directional consistency of its neighbors. Let the direction of drone \(j\) be \(\theta_j(t)\). I define the local directional vector field as
\[
Z_i(t) = \sum_{j \in \Omega_i(t)} e^{i\theta_j(t)}
= r_i(t) e^{i\phi_i(t)},
\]
where \(r_i(t)\) measures the strength of directional agreement in the local neighborhood and \(\phi_i(t)\) is the average direction. If all neighbors move in the same direction, then \(r_i(t) \to 1\). If the directions are randomly scattered, then \(r_i(t) \to 0\).
The migration direction of drone \(i\) is updated as:
\[
\theta_i(t+1) = (1-\lambda_\theta) \theta_i(t) + \lambda_\theta \phi_i(t).
\]
Here \(\lambda_\theta\) controls how strongly the drone follows the local average direction. I set \(\lambda_\theta=1\) in most of my analysis, which means that the drone fully adopts the neighbor-averaged direction at every step.
The migration speed is updated using the following feedback rule:
\[
V_i(t+1) = (1-\lambda_v) V_i(t) + \lambda_v V_{\max} e^{-r_i(t)},
\]
where \(\lambda_v\) is the speed feedback coefficient, and \(\alpha\) in my notation is used interchangeably with \(\lambda_v\). The term \(e^{-r_i(t)}\) implies that when the local directions are highly consistent, the feedback target becomes \(V_{\max} e^{-1}\), which is lower than \(V_{\max}\). In contrast, when directions are inconsistent, the target speed approaches \(V_{\max}\). Thus, the drone slows down in ordered regions and speeds up in disordered regions. This behavior is qualitatively similar to the “slow-to-converge” strategy in UAV formation control.
At every simulation step, the speed update is followed by position update. The direction of motion is given by \(\theta_i(t)\), and the new position is computed by integrating the velocity over the time step.
3.2 Simulation Parameters
The default parameters used in this model are listed in the following table.
| Parameter | Symbol | Default Value |
|---|---|---|
| Population size | \(N\) | \(1000\) |
| Domain length | \(L\) | \(10\) |
| Vision radius | \(R_0\) | \(0.5\) |
| Maximum migration speed | \(V_{\max}\) | varies |
| Speed feedback strength | \(\alpha\) | varies |
| Temptation to defect | \(b\) | varies |
3.3 Effects of Maximum Migration Speed
I first investigate how cooperation changes with the maximum migration speed. The results show that, for every speed-feedback strength, the cooperation level \(C_f\) follows a non-monotonic curve. When \(V_{\max}\) is small, cooperation rises as \(V_{\max}\) increases. There is an optimal value around \(V_{\max}\approx 0.05\) for the baseline setting. Beyond this value, faster migration destroys cooperative clusters and drives the system toward full defection.
The feedback mechanism changes this picture in an important way. When \(V_{\max}\) is large, the model without feedback collapses into full defection quickly. The model with feedback maintains a significant fraction of cooperators even at much higher speeds. The reason is that the feedback mechanism reduces the actual average migration speed below \(V_{\max}\). The drones do not always use the full capability of their maximum speed. Instead, they decelerate when they find a well-aligned neighborhood. This leads to the formation of stable cooperative aggregates.
On the other hand, when \(V_{\max}\) is extremely small, the feedback mechanism can hurt cooperation. Because the feedback further decreases speed, individuals become nearly frozen. Frozen drones cannot escape from defector-rich regions, and cooperators cannot reorganize themselves into compact clusters. This explains why the baseline model without feedback outperforms the feedback model in the very-low-speed regime.
The time evolution of the average speed also supports this conclusion. In the feedback model, the average speed decays from the initial value \(V_{\max}/2\) to a lower steady-state value. In the no-feedback model, the speed immediately jumps to \(V_{\max}\) and stays constant. This distinction is crucial for cooperation because moderate speed allows spatial mixing early in the process, followed by stable clustering in the later phase.
3.4 Effects of Feedback Strength and Temptation
I also examine how the temptation \(b\) interacts with the feedback strength \(\alpha\). In the absence of feedback, cooperation decreases steeply as \(b\) increases and disappears at \(b \approx 1.43\). With strong feedback, the critical threshold moves to a much larger value, \(b \approx 1.56\). Thus, the speed feedback mechanism delays the onset of the defection phase.
When temptation is small, the feedback strength has little effect. When temptation is large, stronger feedback leads to significantly higher cooperation. The reason is that under strong temptation, defectors can easily invade cooperative clusters. Speed feedback allows cooperators to slow down and form denser clusters, which reduces the perimeter exposed to defectors and protects the cluster interior.
3.5 Effects of Interaction Range and Density
The interaction radius \(R_0\) determines the average number of neighbors. I find that very small \(R_0\) makes it difficult for cooperators to support each other. As \(R_0\) increases, cooperation first improves and then declines. There is an optimal interaction range. For small to moderate interaction radii, stronger speed feedback encourages earlier emergence of cooperation. For very large \(R_0\), the population becomes well-mixed and no spatial clustering effect survives; hence the feedback mechanism loses its advantage.
A similar trend is observed when population density is varied. At low density, individuals have few interactions, so cooperation cannot become established. At intermediate density, cooperation reaches its maximum. At high density, the system approaches a well-mixed limit where spatial protection is lost. The speed feedback model is more robust at high density than the baseline model.
4. Adaptive Migration Speed Based on Bush–Mosteller Reinforcement Learning
4.1 Model Description
The second model moves from collective external feedback to individual internal learning. I replace the directional-consistency feedback with a Bush–Mosteller reinforcement-learning mechanism. The drones compare their normalized payoff \(U_i(t)\) with an internal expectation level \(a\). This comparison determines whether the drone feels satisfied or dissatisfied. Satisfaction reinforces the tendency to maintain the current spatial location, while dissatisfaction promotes exploration.
The feedback signal is defined as:
\[
h_i(t) = \tanh\left[\beta\left(U_i(t) – a\right)\right].
\]
In my simulations, I set \(\beta=2\). The hyperbolic tangent function restricts \(h_i(t)\) to the interval \((-1,1)\). When \(h_i(t)>0\), the drone has exceeded its expectation and should reduce its speed. When \(h_i(t)<0\), the drone is dissatisfied and should increase its speed to search for a better environment.
4.2 Speed Update Rule
I use an asymmetric speed update rule that depends on both the current strategy and the learning signal. Let \(s_i(t)=1\) indicate cooperation and \(s_i(t)=0\) indicate defection. The speed at the next time step is:
\[
V_i(t+1)=
\begin{cases}
V_i(t) – h_i(t) V_i(t), & s_i=1,\ h_i(t) \ge 0,\\
V_i(t) – h_i(t) \left(V_{\max}-V_i(t)\right), & s_i=1,\ h_i(t) < 0,\\
V_i(t) – h_i(t) \left(V_{\max}-V_i(t)\right), & s_i=0,\ h_i(t) \ge 0,\\
V_i(t) – h_i(t) V_i(t), & s_i=0,\ h_i(t) < 0.
\end{cases}
\]
This rule creates a behavioral asymmetry between cooperators and defectors. When a satisfied cooperator decelerates, the speed reduction is proportional to its current speed. When a satisfied defector decelerates, the reduction is proportional to the remaining speed capacity. Conversely, when a dissatisfied cooperator accelerates, the increase is proportional to the remaining capacity, while for a dissatisfied defector the increase is proportional to the current speed. This asymmetry leads to different migration patterns for different strategies.
The migration direction in this model is updated by aligning with the local directional average:
\[
\theta_i(t+1) = \phi_i(t).
\]
Thus, the drones align their headings with their neighbors, while the speed is determined by payoff-based learning.
The flow of this learning process is shown below:
| Step | Procedure |
|---|---|
| 1 | Determine neighbors within \(R_0\) |
| 2 | Play weak prisoner’s dilemma with all neighbors |
| 3 | Update strategy with rich-follow rule |
| 4 | Compute payoff satisfaction \(h_i(t)\) |
| 5 | Adjust migration speed using the BM rule |
| 6 | Align direction with local neighbors |
| 7 | Update position and impose periodic boundaries |
4.3 Non-Monotonic Effect of Expected Payoff
One of the most striking findings is that the expected payoff \(a\) has a non-monotonic effect on cooperation. When \(a=0\), all drones are satisfied because all normalized payoffs are non-negative. Hence, every drone tends to decelerate. At low \(V_{\max}\), this excessive deceleration prevents spatial mixing and suppresses cooperation. At moderate and high \(V_{\max}\), the same deceleration helps drones freeze into cooperative clusters, leading to very high cooperation.
When \(a\) is in an intermediate range, such as \(a=0.2\) or \(a=0.5\), the system becomes sensitive to local payoff fluctuations. Many drones are dissatisfied and increase their speed, which destabilizes spatial structures. Cooperation still exists but is less robust. When \(a\) reaches a critical value around \(0.7\), cooperation drops dramatically. The dissatisfaction spreads even inside cooperative clusters, causing them to break apart.
Surprisingly, when \(a=1\), cooperation recovers. At this extreme expectation, almost every drone is dissatisfied because it is rare to receive a payoff of \(1\) while surrounded by defectors. The drones then move at high speed but all of them align to the same direction. In the moving reference frame, relative motion is low. Therefore, spatial reciprocity can be re-established. This phenomenon resembles a global collective motion caused by high aspiration.
4.4 Spatial Clustering and Speed Differentiation
I investigate the micro-structure of the system by measuring the average speed of cooperators and defectors separately. The key result is that cooperators move more slowly than defectors in most parameter regions. Low-speed cooperative clusters act as refuges against exploitation. Defectors, by contrast, move faster and try to exploit these clusters, but they are not able to destroy the dense cooperative core.
The time evolution of the interaction pattern shows that when cooperation is stable, the fraction of cooperative-cooperative links increases, while the fraction of defector-defector links decreases. The average degree of cooperators is much higher than the average degree of defectors. This indicates that cooperators form compact clusters, whereas defectors are expelled to the periphery or scattered in sparse regions.
When the expectation is too high, the speed differentiation disappears. Both cooperators and defectors keep moving, and the spatial structure becomes chaotic. The cooperators lose their density advantage, and the defectors gain the upper hand. This microscopic explanation is consistent with the macroscopic cooperation level observed in the simulations.
4.5 Robustness to Population Size and Interaction Radius
I further test the generality of the BM learning model by varying the average degree \(k\), which can be changed either through the interaction radius \(R_0\) or through the population density. The results show that cooperation as a function of \(k\) follows a universal pattern: it first increases, reaches a peak, and then saturates or slightly decreases. This non-monotonic dependence is robust to whether \(k\) is varied through \(N\) or \(R_0\).
The optimal average degree depends on the expectation level \(a\). For low \(a\), the optimal degree is broader and the cooperation level is very stable. For high \(a\), the optimal interval is narrower and the system becomes more sensitive to structural changes. These findings demonstrate that the cooperation-promoting effect of BM learning is not an artifact of a specific network size.
4.6 Comparison with Alternative Speed Models
To isolate the benefit of the asymmetric BM rule, I compare it with two alternative models. In the symmetric model, both cooperators and defectors use the same speed update rule. In the extreme model, drones either move at \(V_{\max}\) or stop completely, depending on whether the payoff satisfies the expectation.
The symmetric model still promotes cooperation in many cases but performs worse than the asymmetric model when temptation is high. The extreme model is fragile because the sudden “full-speed vs. stop” switching produces strong spatial turbulence. The continuous and asymmetric BM rule is the most robust because it permits fine-grained speed control. Drones can gradually slow down as they approach stable cooperative regions or gradually speed up when they are isolated. This gradual adjustment prevents the violent disruption of cooperative clusters.
The main results of this chapter are summarized in the following table.
| Expected payoff \(a\) | Speed tendency | Cooperation level | Mechanism |
|---|---|---|---|
| \(a=0\) | Strong deceleration | High | Low-speed cooperative clusters |
| \(a=0.2\) | Mixed deceleration | High | Stable clusters with limited exploration |
| \(a=0.5\) | Mixed acceleration | Medium | Frequent cluster fragmentation |
| \(a=0.7\) | Strong acceleration | Low | Destructive spatial mixing |
| \(a=1\) | Global alignment | High | Moving-frame reciprocity |
5. Coupling Directional Consistency with the Bush–Mosteller Algorithm
5.1 Directional Memory and Order Parameters
In the third model, I combine the individual learning mechanism of the BM algorithm with the collective directional-consistency feedback used in the first model. The goal is to create a more complete migration-control framework that accounts for both motion inertia and group synchronization.
I introduce a direction weight \(c\). The migration direction is updated as a weighted combination of the historical direction and the new local field direction:
\[
\theta_i(t+1) = (1-c) \phi_i(t) + c \theta_i(t).
\]
When \(c=0\), the drone completely forgets its previous direction and follows the local average immediately. When \(c=1\), the drone ignores the local field and keeps its original heading. The parameter \(c\) therefore controls the degree of directional memory or inertial persistence.
To adapt the migration speed, I use order parameters. The global order parameter is
\[
r(t) = \frac{1}{N} \left| \sum_{j=1}^{N} e^{i\theta_j(t)} \right|,
\]
which measures the overall directional coherence of the entire swarm. The local order parameter of drone \(i\) is
\[
r_i(t) = \frac{1}{|\Omega_i(t)|} \left| \sum_{j \in \Omega_i(t)} e^{i\theta_j(t)} \right|.
\]
The local order parameter represents the directional agreement in the immediate neighborhood. The feedback signal is now defined by the difference between local and global order:
\[
h_i(t) = \tanh\left[\beta\left(r_i(t) – r(t)\right)\right].
\]
If a drone is in a region that is more ordered than the global average, it tends to decelerate; if it is in a disordered region, it tends to accelerate. The speed is updated using the same asymmetric structure as in the BM model, with \(h_i(t)\) replacing the payoff-difference term.
5.2 Direction Weight and Cooperation
I find that the direction weight \(c\) has a strong nonlinear influence on cooperation. For low maximum speeds, moderate \(c\) values produce the highest cooperation. For high maximum speeds, only high \(c\) values maintain cooperation. This is because high directional memory filters out rapid direction fluctuations and prevents the swarm from being disrupted by random local perturbations.
When \(c\) is very low, drones quickly follow local direction changes, but the chain of local changes propagates in a chaotic way. The frequent direction corrections cause the cooperative clusters to jitter and break. When \(c\) is too high, the drones become insensitive to the local environment. They are unable to move toward cooperative regions or away from defector-rich areas. The optimal \(c\) therefore balances sensitivity and stability.
In the parameter plane of \(c\) and \(V_{\max}\), I observe a distinct “cooperation valley.” This valley occurs when the maximum speed is moderately high and the direction weight is in the intermediate range. In that region, drones move too fast to stabilize clusters but not fast enough to align into a coherent global stream. This is one of the most important findings of the coupled model.
5.3 Emergence of the U-Shaped Sensitivity Response
I also examine the sensitivity coefficient \(\beta\), which controls the steepness of the feedback function. The results reveal a U-shaped relationship between \(\beta\) and cooperation. For small \(\beta\), cooperation is low because the feedback is too weak to generate meaningful speed changes. For intermediate \(\beta\), cooperation is lowest because the system enters a resonant instability: some drones decelerate, while others accelerate, causing spatial fragmentation. For large \(\beta\), the feedback becomes sharply binary. Drones respond decisively and immediately either by stopping or by fleeing, which enables the spatial isolation of defectors.
The high-\(\beta\) regime is analogous to a “freeze-and-flee” strategy. When a drone detects that its local area is less ordered than the global swarm, it immediately speeds up and leaves the region. This prevents defectors from pursuing a single weak target. At the same time, when a drone is in a highly ordered local region, it stops and forms a stable cooperative seed. The sharp environmental response produces a strong spatial defense mechanism.
This U-shaped relation remains valid across different levels of temptation. The cooperation valley deepens as \(b\) increases, and the critical value of \(\beta\) needed to escape the valley also increases. In harsh environments, only highly responsive drones can survive.
5.4 Spatiotemporal Self-Organization
At the spatial level, the feedback mechanism creates a phase separation. Snapshots of the system show that cooperators are organized into dense, directionally aligned aggregates. Defectors do not form stable clusters; instead, they are scattered in the space between cooperative aggregates. The same aggregation pattern appears for different values of \(c\), but the quality of the aggregate changes with the direction weight.
When \(c=0.3\), the aggregates are numerous but small. When \(c=0.5\), the aggregates become larger and more stable. When \(c=1\), the aggregates become rigid and cannot adapt to the changing environment. Thus, there is a sweet spot where the swarms maintain both spatial compactness and directional flexibility.
The coupling of directional consistency and reinforcement learning has a profound effect on the relationship between individual behavior and collective structure. The drones do not need a global controller to determine a common heading. Instead, local feedback interactions make the global order parameter emerge from the bottom up. Speed control then uses this emerging order to decide when to commit to a location and when to leave.
6. Discussion and Implications
The results of my study have several important implications for unmanned aerial vehicles and multi-agent systems in general.
First, the migration rate is not a neutral background parameter. It is an active variable that strongly influences the evolutionary fate of cooperation. The same maximum speed can lead to completely different outcomes depending on how the speed is regulated.
Second, feedback mechanisms must be matched to the spatial and temporal scales of the system. Feedback that is too weak cannot overcome the disruptive effect of fast movement. Feedback that is too strong may freeze the system into a disordered state. My models show that a well-designed speed feedback rule can significantly expand the parameter region where cooperation survives.
Third, individual learning and collective coordination should not be modeled separately. The BM algorithm gives each drone the ability to update its behavior according to past experience. The directional-order feedback gives the system a way to produce a self-organized consensus. When both mechanisms operate together, the swarm can achieve a robust, adaptive, and self-stabilizing cooperative state.
In the context of UAV swarm missions, my models suggest practical control strategies. In a mission, if local alignment is high, the swarm should slow down to maintain formation and protect cooperative interactions. If local alignment is low, some drones should increase speed to re-position. Payoff-based learning can be used to balance individual benefit and group efficiency. The expectation level of each drone is an important design parameter; setting it too high or too low has different consequences.
My findings also provide insights for robust cooperative control. The comparative robustness of the coupled model suggests that unmanned aerial vehicles should be equipped with adaptive speed controllers that use both local sensor information and internal mission-evaluation signals. Such controllers can improve the resilience of UAV swarms to unexpected disturbances, communication failures, and environmental changes.
7. Summary and Future Directions
In this article, I have studied the adaptive decision-making models and co-evolutionary mechanisms of migration rate for unmanned aerial vehicles. I have proposed three progressively more complete models:
(1) A collective feedback model in which the speed of each drone responds to the directional consistency of its neighbors;
(2) An individual learning model in which the speed is updated by a Bush–Mosteller reinforcement-learning mechanism driven by payoff satisfaction;
(3) A coupled model that combines directional memory with local-global order-parameter feedback.
Each model reveals a unique aspect of the relationship between speed and cooperation. The first model demonstrates that external spatial feedback can stabilize cooperative structures in dynamic environments. The second model demonstrates that internal payoff expectations can generate different migration patterns, leading to either cooperative or defective outcomes. The third model shows that the coupling of directional inertia and speed sensitivity is necessary for robust cooperation under extreme conditions.
Future research should consider more realistic constraints related to unmanned aerial vehicles, such as three-dimensional airspace, energy consumption, communication delays, obstacle avoidance, and heterogeneous mission requirements. I also plan to extend the framework to multilayer networks and to multi-task public goods games. Another promising direction is to introduce neural-network-based reinforcement learning so that the speed control can be learned online during a mission. The current framework provides a solid mathematical foundation for these extensions.
Overall, my research bridges the gap between microscopic decision-making and macroscopic ecological evolution in UAV swarms. It shows that speed is not simply a physical quantity; it is a strategic choice. Dynamic regulation of migration speed is an effective mechanism for promoting cooperation and maintaining the order of multi-agent systems.