In my research I investigate how the individual migration behaviour of a single UAV drone can reshape the collective fate of an entire swarm. A swarm of UAV drone platforms is a canonical collective-intelligence system: it has no central controller, it relies on distributed local interactions, and its operational quality depends on how each UAV drone modulates its own motion. Migration is not merely a displacement in space. For a UAV drone it combines speed regulation with directional reorientation, and it therefore encodes the platform’s capacity for environmental adaptation, information exchange and task responsiveness. I build an evolutionary-game framework in which three complementary mechanisms determine the migration rate of every UAV drone: an external speed feedback driven by neighbour heading consistency, an internal payoff-driven Bush–Mosteller reinforcement rule, and a coupled scheme that merges directional memory with rate sensitivity. Across all three models I ask the same question: how does the adaptive control of the migration rate of a UAV drone determine whether cooperation emerges, persists, or collapses?
Background and Motivation
Small UAV drone platforms are cheap, flexible and multifunctional, which makes them attractive for reconnaissance, target support, relay communication and disaster assessment. As mission complexity and environmental uncertainty grow, a single platform becomes insufficient, and the development trajectory therefore moves from monolithic platforms toward swarm-level and intelligent architectures. A UAV drone swarm achieves information sharing, functional complementarity and dynamic coordination through multi-platform cooperation, which improves both operational efficiency and system robustness.
What distinguishes a UAV drone swarm from a centrally commanded formation is that its autonomous coordination arises from a decentralised structure and a very large number of local interactions. To understand the emergence of coordinated swarm behaviour I therefore need to model several layers simultaneously. From the perspective of individual decision making I examine how a UAV drone evaluates payoff, acquires information and updates its policy. From the perspective of group structure I examine how the interaction topology, local feedback and information flow shape cooperation. From the perspective of dynamic regulation I examine how migration, speed adjustment and directional reconfiguration respond to external perturbations and to internal payoff variation, producing spatial aggregation, dispersion or recombination. The migration rate sits precisely at the intersection of these layers, because it determines the interaction range of a UAV drone, the efficiency of information transfer, and the stability of local cooperative structures.
Most classical evolutionary-game studies treat the migration speed as a fixed constant, or they describe migration through the single dimension of directional alignment. Such simplifications can be justified in a static environment, but they fail in strongly time-varying and uncertain missions. A real UAV drone continuously corrects its flight direction according to its own state, its neighbours and its accumulated payoff, and it simultaneously controls its speed to achieve efficient spatial coordination. Migration is thus the bridge that connects microscopic payoff optimisation with macroscopic structural evolution.
Among the two components of migration, direction determines the organisational and communication topology of the swarm, whereas speed determines the intensity and the timeliness with which an individual UAV drone responds to external stimuli and internal payoff variation. Speed is a trade-off between payoff, risk and efficiency, and it is regulated by a learning-like feedback process. I therefore treat the migration rate as the central decision variable of a UAV drone and construct three models that isolate and then combine the external and internal control channels.
Mapping UAV Drone Cooperation onto Evolutionary Games
I map the operational vocabulary of a UAV drone swarm onto the formal vocabulary of the spatial prisoner’s dilemma. The correspondence is summarised in the table below, and it guides every modelling decision that follows.
| UAV drone swarm concept | Evolutionary game concept |
|---|---|
| Swarm of cooperating UAV drone platforms | Structured population |
| Single UAV drone | Individual agent |
| Platforms jointly performing a task | Players |
| Sharing energy, bandwidth or sensing coverage | Cooperation |
| Retaining resources for the private mission | Defection |
| Heading of the UAV drone | Migration direction |
| Speed of the UAV drone | Migration rate |
| Survival gain from selfish operation | Temptation to defect |
| Resources captured inside the local neighbourhood | Payoff |
| Iterative utility-based policy revision | Strategy updating |
| Shifting proportion of strategies in the swarm | Evolution |
| Converged strategy distribution | Evolutionary stable state |
| Density of cooperative UAV drone platforms | Cooperation level |
For each pairwise encounter I use the weak prisoner’s dilemma payoff
$$P_{i,j} = s_i s_j + b\, s_i (1 – s_j),$$
where $s_i = 1$ denotes cooperation and $s_i = 0$ denotes defection, and $b$ is the temptation to defect. The resulting payoff matrix is
| Player 1 \ Player 2 | Cooperation | Defection |
|---|---|---|
| Cooperation | $(1,1)$ | $(0,b)$ |
| Defection | $(b,0)$ | $(0,0)$ |
Because different UAV drone platforms perceive different numbers of neighbours, I normalise the accumulated payoff by the local neighbourhood size,
$$U_i(t) = \frac{1}{|\Omega_i(t)|} \sum_{j \in \Omega_i(t)} P_{i,j}(t),$$
so that a densely connected UAV drone is not advantaged purely by having more interaction partners. Strategy revision follows a strategy-following-the-best rule: the focal UAV drone identifies the neighbour with the highest normalised payoff,
$$j^{\star} = \arg\max_{j \in \Omega_i} U_j(t),$$
and adopts that strategy whenever $U_{j^{\star}}(t) > U_i(t)$; otherwise the platform retains its current policy. All updates are synchronous.
Continuous-Space Model of Interacting UAV Drones
Fixed graph topologies cannot represent the fact that the neighbourhood of a flying UAV drone changes as it moves. I therefore place every platform in a two-dimensional continuous plane of side length $L$ with periodic boundary conditions, so that a UAV drone leaving one edge re-enters from the opposite edge. The position of platform $i$ at time $t$ is
$$\mathbf{p}_i(t) = \bigl(x_i(t),\, y_i(t)\bigr),$$
and platform $j$ belongs to the neighbourhood of platform $i$ whenever
$$d_{i,j}(t) = \sqrt{\bigl(x_i – x_j\bigr)^2 + \bigl(y_i – y_j\bigr)^2} < R_0,$$
where $R_0$ is the sensing radius of the UAV drone. The neighbourhood relation is recomputed at every generation because the positions change. With density $\rho = N/L^2$ the expected number of neighbours of a UAV drone is
$$\bar{k} = \pi R_0^2 \rho = \frac{\pi R_0^2 N}{L^2}.$$
For the reference configuration $N = 1000$, $L = 10$, $R_0 = 0.5$ this gives $\bar{k} \approx 7.85$, which I use as a structural reference value throughout the study.
Positions are advanced according to
$$x_i(t+1) = x_i(t) + V_i(t)\cos\theta_i(t)\,\Delta t, \qquad y_i(t+1) = y_i(t) + V_i(t)\sin\theta_i(t)\,\Delta t,$$
with $\Delta t = 1$. The three models that follow differ only in how the migration rate $V_i(t)$ and the heading $\theta_i(t)$ of each UAV drone are determined. Unless stated otherwise I use $N = 1000$, $L = 10$, $R_0 = 0.5$, a steady state detected after 500 generations without change in the cooperation level, results averaged over the final 500 generations, and 50 independent repetitions per parameter set.
Model I: Direction-Consistency Speed Feedback
The first mechanism treats the migration rate of a UAV drone as a quantity regulated by the local heading consensus of its neighbours. I define the local complex order parameter of platform $i$ as
$$Z_i(t) = r_i(t)\, e^{\mathrm{i}\varphi_i(t)} = \frac{1}{|\Omega_i(t)|}\sum_{j \in \Omega_i(t)} e^{\mathrm{i}\theta_j(t)},$$
so that $r_i(t) \in [0,1]$ measures the degree of heading uniformity inside the neighbourhood and $\varphi_i(t)$ is the mean heading perceived by the UAV drone. The rate update is
$$V_i(t+1) = (1-\lambda_v)V_i(t) + \lambda_v V_{\max} e^{-\alpha\, r_i(t)},$$
and the heading update is
$$\theta_i(t+1) = (1-\lambda_\theta)\theta_i(t) + \lambda_\theta \varphi_i(t).$$
Here $\alpha \in [0,1]$ is the feedback strength, $\lambda_v$ and $\lambda_\theta$ are inertia weights, and $V_{\max}$ is the maximum achievable speed of the UAV drone. When the neighbours of a UAV drone move in nearly the same direction, $r_i \to 1$, the exponential factor becomes small and the platform decelerates to preserve its advantageous neighbourhood. When the local headings disagree, $r_i \to 0$, the platform accelerates to search for a better interaction environment.
Response of Cooperation to Maximum Speed
I first vary the maximum migration rate and the feedback strength. The cooperation level $f_C$ is non-monotonic in $V_{\max}$: it rises slowly at very low speeds, reaches a maximum near $V_{\max} \approx 0.05$, and then decays rapidly. Without feedback the cooperation level collapses completely once $V_{\max} \gtrsim 0.52$. With $\alpha = 1$ cooperation survives up to $V_{\max} = 1$, which shows that the feedback mechanism substantially prolongs the resilience of cooperative clusters.
| Regime | Behaviour without feedback ($\alpha = 0$) | Behaviour with feedback ($\alpha = 1$) |
|---|---|---|
| $V_{\max} < 0.05$ | Slightly higher cooperation | Further deceleration suppresses clustering |
| $0.05 \le V_{\max} \le 0.2$ | Cooperation maintained | Cooperation maintained, slightly lower |
| $0.2 < V_{\max} < 0.52$ | Rapid decay of cooperation | Cooperation sustained at intermediate level |
| $V_{\max} \ge 0.52$ | Complete collapse to defection | Cooperation still present |
I explain this by tracking the swarm-averaged migration rate
$$\bar{V}(t) = \frac{1}{N}\sum_{i=1}^{N} V_i(t).$$
Under the feedback rule the average speed does not remain at its initial value $V_{\max}/2$; it decays during evolution and converges to a lower dynamic equilibrium. This adaptive deceleration indicates that a cooperative UAV drone prefers to reduce its own mobility in order to preserve a favourable neighbourhood structure. In the absence of feedback the average speed jumps immediately to $V_{\max}$ and remains fixed, so the spatial arrangement of the swarm is continuously scrambled and stable policy clusters cannot form.
Interaction with the Temptation to Defect
The benefit of speed feedback is strongly context dependent. At low temptation, $b = 1.2$, the curves for different feedback strengths almost coincide, which means that cooperation is already robust and the feedback has little marginal effect. As $b$ grows, the advantage of large $\alpha$ becomes pronounced. The critical temptation at which the swarm falls into full defection moves from about $b \approx 1.43$ when $\alpha = 0$ to about $b \approx 1.56$ when $\alpha = 1$. In other words, the feedback delays the collapse of cooperation and shifts the phase boundary of the UAV drone swarm toward stronger temptation.
| Temptation $b$ | Cooperation at $\alpha = 0$ | Cooperation at $\alpha = 0.5$ | Cooperation at $\alpha = 1$ |
|---|---|---|---|
| 1.2 | High | High | High |
| 1.3 | Moderate | High | High |
| 1.4 | Low | Moderate | High |
| 1.5 | Near zero | Low | Moderate |
| 1.6 | Zero | Zero | Zero |
Sensing Radius and Density
The sensing radius $R_0$ controls how many partners a UAV drone can evaluate. Very small radii prevent cooperative clusters from forming because an isolated platform is easily exploited. For intermediate radii roughly between $0.2$ and $0.6$ the cooperation level reaches a plateau or a maximum, and larger feedback strengths shift the onset of cooperation to smaller $R_0$. Beyond $R_0 \approx 0.6$ the cooperation level declines regardless of the feedback strength, because excessive mixing destroys the local structure that protects a cooperative UAV drone.
The population density $\rho$ shows the same qualitative structure. At low density the number of encounters is small and cooperation is difficult to sustain, but the feedback mechanism keeps the cooperation level noticeably higher than the no-feedback case. At intermediate density cooperation rises quickly and saturates. At very high density the cooperation level declines again, and the feedback mechanism loses much of its effectiveness. This pattern confirms the existence of an optimal neighbourhood scale for a migrating UAV drone swarm.
Model II: Payoff-Driven Migration under Bush–Mosteller Learning
The second mechanism replaces the external heading feedback with an internal learning rule. Here the migration rate of a UAV drone is regulated by the discrepancy between its realised payoff and an internal aspiration level $a$. I adopt an improved Bush–Mosteller update in which the feedback intensity of platform $i$ is
$$h_i(t) = \tanh\!\bigl[\beta\,\bigl(U_i(t) – a\bigr)\bigr],$$
where $\beta$ controls the sensitivity of the learning rule and $a$ is the aspiration level of the UAV drone. The hyperbolic tangent bounds the feedback inside $(-1,1)$ and prevents extreme payoff differences from producing oscillatory learning. The resulting rate update is asymmetric with respect to strategy and sign:
$$V_i(t+1) = \begin{cases} V_i(t)\bigl[1 – h_i(t)\bigr], & s_i = 1,\ h_i(t) \ge 0,\\[4pt] V_i(t) – \bigl[V_{\max} – V_i(t)\bigr]h_i(t), & s_i = 1,\ h_i(t) < 0,\\[4pt] V_i(t) – \bigl[V_{\max} – V_i(t)\bigr]h_i(t), & s_i = 0,\ h_i(t) \ge 0,\\[4pt] V_i(t)\bigl[1 – h_i(t)\bigr], & s_i = 0,\ h_i(t) < 0. \end{cases}$$
The interpretation is intuitive for a UAV drone: a satisfied platform slows down to consolidate a profitable neighbourhood, whereas an unsatisfied platform speeds up to escape exploitation and search for better partners. The asymmetry between the two coefficients means that cooperators and defectors respond with different effective mobility, because a cooperator multiplies its current speed while a defector adjusts the remaining margin to the maximum.
Aspiration Level as a Control Parameter
The aspiration level $a$ has a strongly nonlinear effect on the fate of the swarm. When $a = 0$ every UAV drone has positive feedback $h_i > 0$ and therefore decelerates. The swarm converges to a slow, densely clustered state, cooperation remains high over a broad range of $V_{\max}$, and the standard deviation of the cooperation level is extremely small, indicating a robust cooperative phase. When the aspiration level is raised to intermediate values the situation changes: a substantial fraction of platforms becomes dissatisfied, accelerates, and destroys the spatial reciprocity that sustains cooperation.
| Aspiration $a$ | Dominant feedback | Resulting migration behaviour | Cooperation outcome |
|---|---|---|---|
| $0$ | Universally positive | Collective deceleration | High and stable |
| $0.2$ | Mostly positive | Mild deceleration | High at low $V_{\max}$ |
| $0.5$ | Mixed | Partial acceleration | Declining with $b$ |
| $0.7$ | Mostly negative | Persistent migration | Lowest, early collapse |
| $1.0$ | Nearly always negative | Continuous co-motion | Recovered to moderate level |
The collapse thresholds confirm this picture. With $a = 0$ cooperation dominates up to $b \approx 1.37$ and survives even beyond $b = 1.6$. With $a = 0.2$ and $a = 0.5$ the full-defection point moves forward to roughly $b \approx 1.45$ and $b \approx 1.35$ respectively. With $a = 0.7$ cooperation disappears already at $b \approx 1.22$, which is the lowest tolerance observed in the entire study.
The most surprising result is that a maximal aspiration level restores cooperation. When $a = 1$ almost no UAV drone is ever satisfied, so the entire swarm migrates continuously. Because the directional update aligns all headings, the swarm settles into a co-moving frame in which the relative spatial structure is effectively frozen. Within that moving reference frame the cooperative UAV drone platforms can again form protected clusters, and the cooperation level rises above the value observed at $a = 0.7$, becoming comparable to the fixed-speed benchmark. This is the same escape mechanism that has been reported for leaving-based migration strategies.
Microscopic Organisation of the Swarm
To expose the mechanism I measure the normalised average degrees of cooperators and defectors, denoted $k_C/\bar{k}$ and $k_D/\bar{k}$, together with the homophilic pair counts $g_C = l_C/l$ and $g_D = l_D/l$. At low aspiration and moderate maximum speed the cooperative UAV drone platforms aggregate into dense clusters with $k_C > \bar{k}$, while most defectors remain sparsely connected. Although defection spreads during the early generations, the spatial segregation that develops later stabilises cooperation.
| Observable | Interpretation for the UAV drone swarm |
|---|---|
| $k_C/\bar{k} > 1$ | Cooperative platforms are more connected than average |
| $k_D/\bar{k} < 1$ | Defectors are dispersed and locally isolated |
| $g_C \to 1$ | Cooperator–cooperator encounters dominate |
| $g_D \to 0$ | Defector–defector encounters become rare |
| $\bar{V}_C < \bar{V}_D$ | Cooperative platforms migrate more slowly |
The average migration speed of the two subpopulations reveals the core mechanism. Cooperators consistently move more slowly than defectors, because a cooperating UAV drone that is surrounded by like-minded neighbours receives enough payoff to trigger deceleration. Once a cooperative cluster has formed, its members become nearly stationary, and the cluster behaves as a spatial refuge. Defectors, by contrast, retain higher mobility and continue to search for exploitable partners. When the aspiration level sits in the valley region, the cooperative cluster is not dense enough to provide mutual support and not mobile enough to escape, so it is eventually invaded. At extreme aspiration every platform, including a cooperator, keeps moving, yet the collective alignment produces an effectively static interaction environment in which a cooperative UAV drone can once again build a stable neighbourhood.
Robustness with Respect to Neighbourhood Size
I test whether the observed cooperation levels depend on how the average degree is generated. I either fix $N = 1000$ and vary $R_0$, or fix $R_0 = 0.5$ and vary $N$. Both procedures change the mean neighbourhood size $\bar{k}$ of a UAV drone, and both produce the same qualitative curve: cooperation first increases with $\bar{k}$, then saturates, and finally declines slightly. The existence of an optimal neighbourhood size $\bar{k}^{\star}$ is therefore a structural property of the swarm rather than an artefact of a particular parameter choice. Error bars are large at small $\bar{k}$, where the system fluctuates strongly between cooperation and defection, they shrink at intermediate $\bar{k}$, and they grow again at very large $\bar{k}$ where over-mixing destabilises cooperation.
Comparison with Alternative Update Rules
I construct two control models to verify that the BM rule is responsible for the observed behaviour rather than the mobility itself. In the symmetric-speed model both cooperators and defectors use the same update,
$$V_i(t+1) = \begin{cases} V_i(t) – V_i(t)h_i(t), & h_i(t) \ge 0,\\[4pt] V_i(t) – \bigl[V_{\max} – V_i(t)\bigr]h_i(t), & h_i(t) < 0. \end{cases}$$
In the extreme-feedback model a dissatisfied UAV drone immediately jumps to $V_{\max}$ while a satisfied platform stops completely. The comparison is summarised below.
| Model | Rate rule | Cooperation behaviour | Stability |
|---|---|---|---|
| Improved BM (present study) | Continuous, strategy-dependent | High over a broad parameter range | Most stable |
| Symmetric speed update | Continuous, strategy-independent | Qualitatively similar, lower tolerance | Moderate |
| Extreme feedback | Binary stop-or-maximum | Rapid collapse at high speed | Weak |
| Fixed speed (no learning) | Constant $V_{\max}$ | Rapid collapse at high speed | Weak |
The extreme-feedback model generates a bimodal stop-or-run dynamics that repeatedly breaks the spatial arrangement and prevents a cooperative UAV drone cluster from forming. The continuous BM rule distributes the adjustment over many generations, which gives the swarm time to reorganise and preserves the cooperative structure.
Model III: Coupling Directional Memory with Rate Sensitivity
The third mechanism combines the external and internal channels. The heading of a UAV drone now obeys an inertia law that blends its previous heading with the neighbourhood mean heading,
$$\theta_i(t+1) = (1 – c)\,\theta_i(t) + c\,\varphi_i(t),$$
where $c \in [0,1]$ is the directional weight. The rate feedback is driven by the difference between the local and the global order parameter,
$$h_i(t) = \tanh\!\Bigl[\beta\,\bigl(r_i(t) – R(t)\bigr)\Bigr],$$
with
$$R(t) = \frac{1}{N}\left|\sum_{j=1}^{N} e^{\mathrm{i}\theta_j(t)}\right|, \qquad r_i(t) = \frac{1}{|\Omega_i(t)|}\left|\sum_{j \in \Omega_i(t)} e^{\mathrm{i}\theta_j(t)}\right|.$$
The migration rate then follows the same strategy-dependent branch structure as in the previous model. The interpretation for a UAV drone is that when its local neighbourhood is more aligned than the swarm as a whole, the platform is locally safe and decelerates; when its local alignment is worse than the global average, it accelerates to relocate. The parameter $c$ plays the role of a low-pass filter on the flight trajectory, and $\beta$ plays the role of a rate sensitivity.
Directional Weight and Maximum Speed
The cooperation level depends on $c$ and $V_{\max}$ in a strongly coupled way. At low maximum speed, moderate directional weighting $c \approx 0.5$ produces the best cooperation, and the curves for different $c$ are close together. As $V_{\max}$ grows, a pronounced U-shaped valley appears in the $c$-dependence of $f_C$: intermediate values of $c$ give the worst outcome, because the platform possesses neither the flexibility of small $c$ nor the structural stability of large $c$. With very high maximum speed the cooperation peak on the low-$c$ side disappears entirely, and cooperation can only be maintained in the limit $c \to 1$.
| Maximum speed $V_{\max}$ | Optimal directional weight | Observed pattern |
|---|---|---|
| 0.2 | Any $c$ in a broad range | High cooperation across the whole range |
| 0.3 – 0.4 | $c \approx 0.5$ | Shallow valley at intermediate $c$ |
| 0.5 – 0.6 | $c \approx 0.7$ | Deep U-shaped valley |
| 0.8 – 1.0 | $c \to 1$ | Only strong inertia sustains cooperation |
The optimal directional weight therefore shifts toward larger values as the maximum speed increases. The reason is that the final average migration rate
$$\bar{V} = \frac{1}{N}\sum_{i} V_i$$
decreases monotonically with $c$, and the cooperation level is maximised when $\bar{V}$ is close to the optimal migration scale $V_{\mathrm{best}} \approx 0.05$. When $V_{\max}$ is small, even a modest $c$ keeps $\bar{V}$ near $V_{\mathrm{best}}$. When $V_{\max}$ is large, a small or moderate $c$ allows $\bar{V}$ to drift far above $V_{\mathrm{best}}$, and only a large $c$ restores the rate to the favourable interval. This rate-matching argument explains the systematic rightward shift of the optimal directional weight.
Feedback Intensity and Spatial Locking
The average feedback intensity $\bar{h}$ stays positive and large, typically between $0.7$ and $1.0$, across all configurations, which is precisely why the migration rate is suppressed. At low maximum speed the curves for different $c$ converge quickly to $\bar{h} \approx 1$, so almost every UAV drone becomes locally locked. At high maximum speed the curves fan out and form a staircase: larger $c$ produces higher $\bar{h}$, which in turn produces a smaller average migration rate and a more stable cooperative cluster.
| Directional weight $c$ | Average feedback $\bar{h}$ | Final rate $\bar{V}$ | Cooperation state |
|---|---|---|---|
| 0.3 | Moderate, declining late | High | Cluster gradually torn apart |
| 0.5 | High and stable | Intermediate | Stable cooperative plateau |
| 0.7 | Very high | Low | Robust cooperative cluster |
| 1.0 | Near saturation | Very low | Vector-frozen, low cooperation |
When $c = 1$ the UAV drone preserves its historical heading and ignores the new target direction. The migration rate becomes so low that the platform is effectively frozen in space. Although local alignment persists, the swarm loses the ability to reconfigure, so cooperative structures remain loose and the cooperation level stays low. There is therefore an interior optimum: a moderate directional weight acts as a spatial low-pass filter that admits enough reconfiguration while suppressing destructive scrambling.
Temptation to Defect and Rate Sensitivity
The cooperation level decreases monotonically with the temptation $b$ for every value of $c$, but the curves are not ordered in a simple way. The value $c = 0.5$ produces the highest cooperation over most of the range, whereas $c = 1$ produces the lowest. The cases $c = 0.3$ and $c = 0.5$ cross: for $b < 1.4$ the smaller weight performs better, and for $b > 1.4$ the moderate weight performs better. This crossing reflects the competition between rapid local reconfiguration, which helps under weak temptation, and stable cluster defence, which helps under strong temptation.
The rate sensitivity $\beta$ introduces a further U-shaped dependence. When $\beta$ is small, a UAV drone responds weakly to the gap between local and global alignment, so it cannot cut its interaction with a defector quickly enough, and cooperation collapses early. When $\beta$ is large, the platform reacts decisively at the moment of exploitation, which preserves the cooperative cluster. Increasing $\beta$ from $0.5$ to $2.0$ shifts the full-defection threshold from about $b = 1.42$ to about $b = 1.50$.
| Sensitivity $\beta$ | Response of a single UAV drone | Swarm-level consequence |
|---|---|---|
| 0.4 – 0.6 | Sluggish rate adjustment | Deep cooperation valley, early collapse |
| 1.0 | Moderate adjustment | Intermediate cooperation |
| 1.5 | Fast adjustment | Robust cooperative clusters |
| 2.0 | Decisive adjustment | Highest tolerance of temptation |
Comparative Synthesis
The three mechanisms operate on different information channels but converge on the same physical principle. A cooperative UAV drone survives only when it can reduce its effective migration rate at the right moment, so that its neighbourhood becomes a stable and mutually supportive cluster. The external heading feedback achieves this through local consensus, the internal BM rule achieves it through payoff satisfaction, and the coupled scheme achieves it through the difference between local and global alignment.
| Feature | Model I: speed feedback | Model II: BM learning | Model III: coupled scheme |
|---|---|---|---|
| Information source | Neighbour headings | Own payoff versus aspiration | Both payoff gap and alignment gap |
| Control variable | Feedback strength $\alpha$ | Aspiration $a$, sensitivity $\beta$ | Directional weight $c$, sensitivity $\beta$ |
| Key nonlinearity | Non-monotonic in $V_{\max}$ | Non-monotonic in $a$ | U-shaped in $c$ and in $\beta$ |
| Critical temptation | Extended to $b \approx 1.56$ | Extended to $b \approx 1.60$ | Extended to $b \approx 1.50$ |
| Microscopic signature | Low-speed cooperative clusters | Low-speed, high-density cooperators | Spatially locked cooperative clusters |
| Robustness to density | Optimal $R_0$ interval | Optimal $\bar{k}^{\star}$ | Optimal $c$ shifts with $V_{\max}$ |
The combined picture is that the migration rate of a UAV drone is not a fixed physical parameter but an adaptive decision variable. The swarm succeeds when the rate of each platform is continuously matched to the structural requirements of cooperation. Too slow a UAV drone cannot escape exploitation or explore space; too fast a UAV drone destroys the local correlation that makes reciprocity profitable. The three mechanisms are three different ways of steering the swarm toward the favourable intermediate regime.
Outlook
Several limitations remain and I regard them as directions for further work.
| Direction | Motivation | Planned extension |
|---|---|---|
| Realistic flight constraints | Obstacles, communication delay, energy limits, sensor error | Embed physical-layer constraints into the rate rule |
| Three-dimensional airspace | A real UAV drone swarm operates in layered volumes | Multi-layer and volumetric game models |
| Individual heterogeneity | Platforms differ in preference, priority and risk attitude | Heterogeneous agents with asymmetric aspirations |
| Task-driven payoffs | Mission success is not captured by a single scalar | Multi-objective payoff structures |
| Scalability | Large swarms require efficient computation | Distributed and event-triggered updates |
Introducing obstacles and communication delays would allow the model to describe how a UAV drone reconfigures its rate when its sensing is degraded. Extending the plane to a layered three-dimensional volume would make it possible to study how altitude-dependent interaction ranges affect the stability of cooperative clusters. Allowing heterogeneous platforms would connect the framework to realistic mission planning, in which one UAV drone may prioritise coverage while another prioritises endurance. Finally, replacing the scalar payoff with a multi-objective utility would let the swarm trade off coverage, energy and connectivity simultaneously.
Closing Remarks
I have constructed an evolutionary-game framework for the migration behaviour of a UAV drone swarm and analysed three adaptive decision models for the migration rate. The first model regulates the rate through neighbour heading consistency and shows that an appropriate feedback strength delays the collapse of cooperation and shifts the critical temptation toward larger values. The second model regulates the rate through the discrepancy between realised payoff and an internal aspiration level, and it reveals a strongly nonlinear dependence of cooperation on aspiration: intermediate expectations cause disordered migration, whereas very low and very high expectations promote cooperation through distinctly different dynamical routes. The third model couples directional memory with rate sensitivity and demonstrates that a suitable directional weight triggers feedback-driven deceleration and induces stable cooperative patches, while the rate sensitivity exhibits a clear U-shaped relationship with the cooperation level.
Across all three models the same conclusion emerges. The maintenance of cooperation in a mobile UAV drone swarm depends less on the absolute speed of the platforms than on how each platform adapts its migration rate to its local structural context. The effective coupling of rate regulation and directional optimisation improves the overall adaptability of the swarm, provides a theoretical basis for understanding dynamic stability in time-varying environments, and offers a generalisable modelling and analysis framework for self-organisation and multi-agent cooperative control in collective-intelligence systems.