General Patrol Problem Overview
- General Patrol Problem is a family of formulation models addressing repeated visitation for optimizing surveillance, detection, and coverage across diverse structures such as graphs, perimeters, and metric spaces.
- It incorporates paradigms like adversarial patrol games, min-max latency scheduling, and recurrent structural solutions, offering insights into both theoretical performance and practical deployment.
- Research blends exact algorithms for structured environments with scalable, learning-based methods to adapt to variable network sizes and heterogeneous resource constraints.
Searching arXiv for recent and foundational papers on patrolling formulations, perimeter patrol, periodic patrols, adversarial patrol games, and multi-agent patrol scheduling. arXiv search query: "general patrol problem patrolling perimeter periodic patrol line graph multi-agent latency adversarial visible patroller" The General Patrol Problem denotes a family of patrol design problems in which one or more agents repeatedly visit locations in a perimeter, graph, line, metric space, or network so as to optimize an objective defined over surveillance quality, attack detection, revisit latency, or weighted coverage. Across the literature, the problem appears as a continuous-time Stackelberg patrol game on a closed perimeter, a discrete-time patrolling game on graphs, a min-max latency scheduling problem in metric spaces, a finite-horizon cooperative coverage problem on urban graphs, and a broad graph-based idleness minimization model with node-wise constraints (Lin, 2019, Afshani et al., 2020, Palma-Borda et al., 14 Jan 2025, Mallya et al., 15 Sep 2025). A recent unifying result for non-adversarial graph patrol states that if the General Patrol Problem has a solution, then it has an -optimal recurrent solution, with proportional to a discretization constant and independent of the number of patrol agents and the size of the environment (Mallya et al., 15 Sep 2025).
1. Formal scope and objective classes
A common core across formulations is a patrol state space, a movement model, and an objective defined on repeated visitation or defense. In the broad graph formulation, the environment is a strongly connected directed graph with node weights and edge traversal times, a patrol strategy is an infinite ordered sequence of departure events, and the General Patrol Problem is to minimize
subject to node-wise idleness constraints on a designated subset of nodes, with
This template subsumes Graph Maximum Idleness, Weighted Maximum Idleness, and Graph Average Idleness (Mallya et al., 15 Sep 2025).
In the classical adversarial formulation, a patrolling problem is
where is the set of admissible defender positions, is the set of vulnerable targets, 0 is the defender’s initial node, 1 is the directed move relation, and 2 assigns an attack length to each target. A defender strategy is a history-dependent randomized policy over successors, and the attacker chooses a target and a launch time after observing the defender’s realized path (Brázdil et al., 2015). In the Internet patrolling model, the game is
3
where 4 is the target set, 5 is the intrusion duration, 6 is target importance, and 7 is the per-visit detection probability. There are 8 centrally coordinated software patrollers, time is discrete and infinite-horizon, and the defender’s objective is to maximize the level of protection
9
The optimization target is
0
A separate but closely related line of work formulates patrol quality through revisit time. In the multi-robot min-max latency problem, a schedule consists of continuous trajectories 1 satisfying a unit-speed Lipschitz condition, site latency is the supremum of intervals between consecutive visits, and the objective is
2
or, in the weighted version,
3
For 4, the problem is equivalent to TSP (Afshani et al., 2022, Afshani et al., 2020).
A plausible implication is that “General Patrol Problem” is best understood not as one canonical optimization problem, but as a family of patrol formulations sharing repeated visitation, limited patrol capacity, and worst-case or best-response evaluation.
2. Adversarial patrol games and observability
The adversarial branch of the literature models patrol as a zero-sum or Stackelberg interaction between a defender and an attacker who times attacks against observed or known patrol behavior. In the continuous-time perimeter model, the defender dispatches patrollers from a base on a closed perimeter under a long-run dispatch-rate cap 5, each patroller traverses the entire perimeter exactly once, and an attack of duration 6 is detected independently by each passing patroller with probability 7. If 8 patrol passes occur during the attack interval, then
9
so the unconditional detection probability is
0
The main theorem shows that the optimal value is
1
and, remarkably, that visible patrols need not be worse than invisible patrols: with appropriate probabilistic spacing, the defender achieves the same value in both cases (Lin, 2019).
The paper’s central mechanism is variance control. For a fixed mean 2, the patrol-count distribution maximizing detection is the least-variable one: 3 deterministically if 4 is an integer, and otherwise a two-point distribution on 5 and 6. This yields the “deterministic lattice plus Bernoulli augmentation” schedule on the perimeter, under which every attack window of length 7 sees either 8 or 9 patrols with the correct probabilities (Lin, 2019). This directly contradicts the common intuition that visible patrol must always hurt the defender.
A different informational effect appears in the Uniformed Patroller Game on a star network. There, the attacker observes whether the patroller is present at the chosen target, waits until the patroller leaves, and then chooses a delay 0 before launching an 1-period attack. The paper’s principal structural theorem states that for every 2 and 3, the attacker’s optimal delay is 4 and the patroller’s optimal end-node behavior is 5, meaning the patroller should never remain at the same visited end node for a second consecutive period. In the 6 case, visible presence can reduce thwarted attacks by as much as a factor of four in specific models, and the asymptotic ratio to the non-uniformed value approaches 7 (Alpern et al., 2019). Here visibility is harmful because the attacker exploits local departure information.
Adversarial perimeter patrolling with multiple robots and a full-knowledge opponent has also been analyzed through a Markovian patrol model in which robots move on a discretized perimeter or fence and continue straight with probability 8 or reverse with probability 9. The optimization criterion is
0
where 1 is the penetration-detection probability at segment 2. In the perimeter case, the paper proves that optimal patrol within this Markovian class requires uniform spacing and coordinated motion, and computes the optimal parameter in polynomial time by maximizing the lower envelope of the segment-wise detection functions (Agmon et al., 2014).
At the most general level, adversarial patrolling on arbitrary finite environments admits an optimal defender strategy for every game instance, and for every 3 there exists a finite-memory 4-optimal defender strategy constructible in exponential time. The same work shows that no polynomial-time procedure can compute such strategies in general unless 5 (Brázdil et al., 2015). This places the continuous-time perimeter and star-network models in a larger game-theoretic context: exact sharp solutions exist in highly structured settings, while arbitrary environments require approximation or special structure.
A related heterogeneous-resource Stackelberg model appears in community-participated patrol. There the defender allocates continuous professional patrol effort 6 and integer community patrol assignments 7 under
8
with additive capped coverage
9
The attacker then chooses a best target, and the defender maximizes induced expected utility (Wu et al., 2024). This model shows that patrol observability and attacker best response remain central even when resources are heterogeneous and partially inflexible.
3. Exact and near-exact results on perimeter, line, fence, and periodic graph structures
Highly structured environments have yielded some of the sharpest results in patrol theory. On the perimeter, the one-location reduction in the continuous-time model shows that what matters is the patrol-count law in each attack window, not raw unpredictability. The optimal perimeter policy dispatches according to the one-location mixed schedule and has every patroller circumambulate the perimeter in the same direction at the same constant speed; direction mixing and speed variation do not improve the value so long as each patroller passes each point exactly once and does not reverse direction mid-trip (Lin, 2019).
Periodic graph patrolling gives a complementary discrete-time perspective. For a graph 0, period 1, and attack duration 2, a patrol is a 3-periodic walk. When 4 is even, the value on an arbitrary graph is exactly
5
where 6 is the fractional covering number and 7 is the fractional independence number. An optimal patroller oscillates on edge 8 with probability 9, where 0 is an optimal fractional edge cover, and an optimal attacker chooses a fixed 2-period interval and attacks node 1 with probability 2 (Alpern et al., 2017). For line graphs 3, the paper gives a complete solution for all 4 and all periods 5, with the odd/odd case determined by the threshold 6 and the exact value
7
when both 8 and 9 are odd (Alpern et al., 2017).
Fence patrolling by mobile agents provides another one-dimensional benchmark. For an open fence of length 0, the partition strategy 1 assigns each agent its own subsegment and achieves idle time
2
For a closed fence under unidirectional patrolling, the runners algorithm 3 yields
4
The paper disproves the conjecture that 5 is always optimal, and for open fences gives a constructive improvement over partitioning: for every 6, there exist instances with idle time at most
7
improving the earlier 8 separation (Dumitrescu et al., 2014). The same paper also shows that some circle instances have “useless” agents, whereas on a segment any optimal schedule must use all agents (Dumitrescu et al., 2014).
The one-dimensional adversarial multi-robot model on perimeters and fences complements these results by optimizing the weakest-point detection probability rather than idleness. There, deterministic patrol may yield 9 against a full-knowledge adversary, while randomized Markovian patrol ensures positive detection probability and admits polynomial-time optimization in the structured one-dimensional setting (Agmon et al., 2014). This suggests that frequency-optimal patrol and adversarially robust patrol need not coincide.
A static covering analogue appears in the geometric problem of Guarding a Set of Segments. Given a finite set of straight line segments, one seeks a minimum-cardinality subset of vertices 0 such that
1
This problem is strongly NP-complete even for cubic graph structure, but solvable in 2 time when the segment network is a tree (Brimkov, 2010). Although this is a static placement problem rather than a moving-patrol schedule, it isolates the role of network topology in patrol resource allocation: cyclic street systems are computationally hard, tree structures are not.
4. Persistent monitoring, weighted latency, recurrent structure, and limited memory
Outside adversarial timing games, a major patrol objective is to minimize the worst revisit time of sites. In the min-max latency multi-robot patrolling problem on a metric space, a cyclic solution partitions the sites into groups, assigns robots to groups, and places the assigned robots evenly on a TSP tour of each group. If the TSP tour of group 3 has length 4 and 5 robots are assigned to it, the group latency is 6. The paper proves that an optimal cyclic solution is a
7
-approximation of the unrestricted optimum, and for 8 it is optimal overall. It also reduces approximation of the best cyclic patrol to approximation of TSP with only a 9 loss in approximation factor and a runtime overhead 00 for fixed 01 (Afshani et al., 2022).
A complementary weighted formulation studies patrol scheduling with objective
02
where 03 is the latency of site 04 and 05 its weight. In general metrics, the paper gives a polynomial-time
06
-approximation. In one dimension, the equal-weight case is exactly solvable: there exists an optimal schedule in which each robot follows a zigzag schedule and the intervals are disjoint. For weighted sites on a line, the paper gives a 07-approximation when the number of robots is constant (Afshani et al., 2020). These results establish a rigorous approximation baseline for persistent monitoring versions of the patrol problem.
Recent work on the formal General Patrol Problem sharpens the structural picture. For graph-based patrolling with idleness-based objectives, every feasible patrol strategy admits an 08-approximate recurrent patrol strategy, and if the problem has a solution, then it has an 09-optimal recurrent solution. The approximation factor is proportional to a discretization constant 10, and the paper emphasizes that this dependence is independent of the number of agents and of the environment size (Mallya et al., 15 Sep 2025). The proof proceeds by time discretization to a strategy 11, then by extracting a repeated global state and repeating the intervening segment indefinitely. This establishes recurrence as a general near-optimal structural class rather than a heuristic restriction.
The role of structural regularity is even sharper on finite grids with minimal memory. For a 12-dimensional grid graph, the paper gives an exact characterization of when oblivious (13-bit) patrol is possible with sensing range 14, and proves that 15 bit of persistent memory with visibility 16 suffices to patrol every finite 17-dimensional grid graph. The positive result is constructive, and every vertex is visited within at most 18 steps (Amir et al., 2023). This shows that in highly regular environments, patrol feasibility can depend more on tiny amounts of state than on large sensing range.
Taken together, these works indicate that recurrent, cyclic, or periodic patrol structure is often close to optimal, sometimes exactly optimal, but not universally so. This suggests a productive distinction between structural sufficiency results and unrestricted optimal control: recurrent schedules are broadly powerful, yet the literature also contains explicit examples where optimal patrol is nontrivial, non-memoryless, or not reducible to simple partitioning.
5. Scalable synthesis, guarding, resource heterogeneity, and adaptation to attrition
A distinct branch of the literature addresses large-scale patrol synthesis under implementation constraints. In the fully connected Internet patrolling model, the defender uses centrally coordinated software patrollers and seeks to maximize protection against an attacker who observes the committed strategy and past patrols. The main algorithmic ideas are modular strategies, decomposition into basic sets, and a compositional synthesis principle that assigns each basic set either 19 or 20 patrollers with suitable probabilities. The construction is provably optimal when all basic sets are full, and near-optimal in general, with the paper reporting that it can quickly construct (sub)optimal strategies for instances with hundreds of millions of targets (Brázdil et al., 2018). This is one of the clearest demonstrations that the general patrol problem can sometimes be reduced from enormous target sets to a much smaller type-based optimization.
Attrition-aware multi-agent patrolling introduces a different systems constraint: agents may be lost during execution, and adaptation should require minimal communication. The environment is a graph 21 with patrol points, each agent starts from an origin node, and patrol quality is average idleness. The Adaptive Heuristic-based Patrolling Algorithm uses Voronoi partitioning to assign patrol points and nearest-neighbor TSP to order visits within each partition. For homogeneous-speed agents, the paper proves that loss of a single agent changes only the allocations of its neighboring agents under the Voronoi heuristic, and it gives approximation guarantees inherited from multi-depot routing. Empirically, the method adapts to agent loss while requiring only a single message in case of attrition (Goeckner et al., 2023). This makes explicit a design criterion that is often implicit in patrol theory: adaptation cost and communication burden can be as important as nominal patrol quality.
Heterogeneous defender resources appear in the community-participated patrol model, where the defender allocates continuous professional effort and integer community assignments with additive capped coverage
22
The paper gives both an additive-approximation algorithm, Two-Dimensional Binary Search, and an exact polynomial-time Hybrid Waterfilling algorithm with runtime
23
The latter combines greedy placement of indivisible community patrollers, water-filling over flexible professional effort, and discrete swap operations to minimize wasted community coverage (Wu et al., 2024). A further tractability boundary is especially informative: if community effectiveness becomes villager-specific rather than target-specific, the problem becomes NP-hard (Wu et al., 2024). This is a sharp statement about how quickly mixed patrol allocation becomes combinatorial once indivisible resources are heterogeneous individually.
Static guarding of street networks adds a topological allocation perspective. Guarding a Set of Segments is strongly NP-complete in general but solvable in polynomial time on trees (Brimkov, 2010). This suggests a broader pattern already visible in dynamic patrol: topology, divisibility of resources, and communication assumptions are not secondary modeling details; they determine whether exact synthesis, near-optimal decomposition, or only heuristic planning is realistic.
6. Data-driven and learning-based patrol planning
Recent work increasingly treats patrol planning as a sequential decision problem learned from data rather than derived analytically from a small symbolic model. In energy-aware and fault-tolerant deep reinforcement learning for multi-agent patrolling, the environment is a grid graph with charging stations, agents move with actions 24, and the patrol objective is to minimize both average and time-averaged worst-case idleness, measured through
25
The reward combines patrol quality and battery behavior,
26
with difference rewards for credit assignment and a shared-policy MAPPO architecture under centralized training and decentralized execution. The system is designed to tolerate agent failures and supplementary agents, and in simulation it achieves greater than 27 recharge success while outperforming a Conscientious Reactive heuristic on all four maps and all tested team sizes from 28 to 29 (Tong et al., 2022).
A finite-horizon graph-based crime-surveillance model formulates coordinated patrol routing as a Dec-POMDP on an undirected graph
30
where 31 is a node-importance function derived from crime data. The reward discounts repeated short-term revisits and adds exploration bonuses, and the paper evaluates patrol quality with a new coverage index over the top-32 highest-value nodes. Using VDPPO, the learned routes achieve more than 33 coverage of the 34 of graph nodes with the highest crime incidence, and 35 for 36 of these nodes (Palma-Borda et al., 14 Jan 2025). Unlike min-max idleness models, this formulation is explicitly finite-horizon and priority-weighted: complete repeated coverage is not the goal.
A related multi-officer routing model combines crime data, check-ins, POIs, and incident response data. The environment is represented as a graph, node benefit is
37
and the multi-officer objective is
38
subject to time-budget and non-overlap constraints. Route optimization is done with hotspot-aware G-LERK encoding plus Genetic Algorithm or Cuckoo Search, and experiments show that guided methods outperform unguided variants especially when the number of officers is limited (Rumi et al., 2020). This line of work shifts the patrol problem from steady-state revisit control to predictive, reward-maximizing route design under a day-shift horizon.
These learning-based formulations broaden the operational meaning of the General Patrol Problem. They retain the classical concerns of coverage, revisit regularity, and attacker response, but combine them with battery management, recharge logistics, incident response, heterogeneous observations, and real urban data. This suggests that modern patrol research is increasingly bifurcated: one branch seeks exact structural theory, and another seeks scalable policies under realistic state, data, and execution constraints.