A Game-Theoretic Approach for the Robust Daily Aircraft Routing Problem

In the operation of airlines, the most important link is determining the route and scheduling of aircraft. Te key to this problem is to input the fight segment and aircraft type and fnally determine all fight segments for each aircraft. In this paper, we focus on fnding feasible, robust scheduling for various uncertainties in the fight process. Tis paper presents a new robust integer mathematical model based on game theory that considers daily aircraft routing. Ten, in order to fnd the suboptimal solution to a large-scale integer programming problem in a limited amount of time, a heuristic algorithm integrating a column generation algorithm and variable domain search is introduced. In addition, we use the data of a Chinese airline to verify, and the experimental results show that the model proposed by us is more robust than the model in general.


Introduction
With the increasingly ferce competition in the aviation industry, an efective decision is very important to the proftability of airlines. However, designing an aviation network is a very complex task. Terefore, many researchers generally divide the design of the entire aviation network into a series of subproblems, including route trafc volume prediction, fight planning, aircraft assignment, crew pairing, and crew rostering. Te route trafc volume forecast is based on the trafc volume of the national air transport market, the route market, and the economic situation of a region. Flight planning is to generate fight schedules based on the previously predicted route trafc volume. Aircraft scheduling is to generate an aircraft scheduling plan based on the aircraft type selected for the fight plan and other relevant constraints. Duty scheduling is to generate a fight segment pairing that meets the requirements of a fight crew's one-day mission according to the required fight time and laws and regulations in the fight plan, that is, the duty schedule. Te crew rostering generates the fight schedule in accordance with laws and regulations based on the generated fight pairings and crew members. In this paper, we focus on the aircraft assignment phase.
Although there are more and more routes and the route structure is more and more complex due to the restrictions on airspace resources, airport facilities, and the various conditions of airlines, coupled with weather reasons, fight delays and even fight cancellations are becoming more and more frequent, which has gradually become a major problem in the development of civil aviation. According to the data of the Civil Aviation Administration of China, with the increase in the number of fights and routes, the normal rate of fights has shown a downward trend. Flight plans with strong anti-interference can not only reduce fight delay to a certain extent but also save some costs for airlines to a certain extent. Whether the fight plan is antijamming is mainly refected in its robustness. If a fight is delayed for some reason, it is likely to cause successive delays of subsequent fights. Terefore, it is of great practical signifcance and value to develop a reasonable aircraft scheduling plan based on the actual operation to reduce the airline delay rate. At present, civil aviation usually divides fight plans into two categories according to their functions during operation. On the one hand, it is an "advance strategy," that is, for the prepared fight plan, the transit time should be adjusted appropriately according to the actual operation to absorb the departure delay caused by previous fights as much as possible so as to improve the robustness of fight plans. On the other hand, it is an "after the fact strategy," which is to take certain measures and schemes in time after fight delays, such as aircraft transfer, aircraft exchange, or even cancellation of fights, so as to avoid causing larger delays. In this paper, we mainly solve a daily aircraft route problem (DARP), including route planning and route assignment. Te advanced strategy is mainly used to adjust fight cost and fight plan robustness and try to improve the robustness of the fight plan on the basis of cost control.

Literature Review.
In fact, the daily fight route problem is a classical routing problem. Terefore, in this section, we frst review the literature on daily aircraft route problems, then review the application of game theory, and fnally review the application of meta-heuristic algorithms to engineering problems.
Tis problem was frst proposed in 1971 by Levin [1], who modeled the problem as a 0-1 programming model and solved large-scale instances using Land-and-Doig techniques. Desaulniers et al. [2] model the problem as a set of daily aircraft route allocation models and multicommodity fow models to improve airline efciency and reduce costs, and solve both models using column generation and Dantzig-Wolfe decomposition, respectively. To solve this model, an iterative solution method was proposed. In order to solve the feet allocation problem and the aircraft routing problem at the same time, Barnhart et al. [3] proposed a model and solution method to solve the two problems simultaneously. Sosnowska [4] successfully solved the fight data of a medium-sized company using a method based on a simulated annealing algorithm and GRASP. Mercier and Soumis [5] synthesized the aircraft routing problem and the crew scheduling problem, proposed a basic ensemble model, and solved this large-scale optimization model based on the Benders decomposition. Weide et al. [6] integrated the aircraft routing problem and the crew scheduling problem and established a large-scale 0-1 programming model considering the robustness of the problem. Jamili [7] proposed a mixed-integer programming model that integrates aircraft robustness and a hybrid heuristic that yields a more efcient solution. Kenan et al. [8] propose a two-stage planning model based on the uncertainty of demand and fare and solve the model using the method of sample average approximation. Considering the uncertainty of the data, Cadarso and de Celis [9] propose a large-scale mathematical model, which is solved by the Benders decomposition approach. Cui et al. [10] proposed 3 models, improved the VNS algorithm, compared the experimental results with commercial solvers, and verifed the efectiveness of the algorithm. Si et al. [11] proposed a multicommodity fow model and an arc fow model, improved the column generation algorithm, and signifcantly reduced the running time. Xu et al. [12] proposed a mixed integer programming model considering the infuence of propagation delay and fight retiming decisions and proposed column generation as well as variable neighborhood search for a solution. Şafak et al. [13] proposed a new two-stage stochastic decision-dependency programming model for airline network expansion. In fact, the daily fight route problem for aircraft is a routing problem. For routing problems, many studies were conducted during the new epidemic, especially on the supply chain problem [14].
Based on the previous work, the main purpose of this paper is to fnd a schedule that ensures the lowest cost incurred when some fights are delayed. Tis paper mainly develops a new method to deal with fight delays caused by various uncertain factors to maintain the stability of the system. Based on the abovementioned remarks, we establish a robust optimization model based on game theory.
Game theory is a commonly used optimization method. Game theory has been used in many felds, such as transportation and power systems. Lima et al. [15] constructed a cooperative game theory model based on the cooperative game theory framework, found a loss allocation solution, and, compared with other traditional loss allocation methods in transmission power systems. In [16], a beneft distribution method is proposed using a cooperative game, and the result shows that power producers will get more profts by cooperating in competition. In [17], to ensure sufcient power generation, a game theory-based approach to power system reserves is proposed, with planners and nature as two players. Compared with traditional methods, the proposed game theory method has better robustness and higher efciency. In [18,19], based on a game theory framework, a robust railway transportation network is designed for line failures.
At the same time, because it is very difcult to solve the daily fight route problem with an accurate algorithm, this paper combines the column generation algorithm and a meta-heuristic algorithm to solve it. Te meta-heuristic algorithm can obtain a feasible solution close to the optimal solution in a limited time and has been widely used in engineering problems. For example, Alinaghian and Goli [20] used an improved harmonious search algorithm to solve the confguration problem of temporary medical centers in rural areas in crisis. Yang et al. [21] calibrated surface turbulence-related parameters in the source area of the Yellow River using a particle swarm optimization algorithm. Yang et al. [22] used a particle swarm optimization algorithm to efectively calibrate freezing and thawing-related parameters and improve simulation accuracy.
Based on the abovementioned literature review, according to the author's current knowledge, although some scholars have studied integrated aircraft routing and scheduling before, they have not yet used the game theory method for research. As can be seen from the previous literature on game theory, game theory is an efective way to deal with uncertain problems. At the same time, a metaheuristic algorithm is an efective method to solve these complex models.

1.2.
Contribution of the Paper. Tis paper presents a new model of integrated aircraft routing and scheduling that considers feet assignment. In addition, there are many uncertain factors that cause fight delays, including weather and air trafc control. Terefore, we introduce a robust method based on game theory into the model. Finally, because the model is difcult to fnd the optimal solution in a limited time, a method integrating column generation and variable neighborhood search is proposed to solve this model.

Outline.
Te main parts of this paper are as follows: In Section 2, the problem is described and a new integer programming model is established. In Section 3, we propose a method to solve large-scale integer programming by integrating column generation and variable neighborhood search. Sections 4 and 5 are numerical experiments and conclusions, respectively.

Problem Definition
DARP requires aircraft to be allocated to each fight segment to minimize cost. Te takeof time and landing time of each fight segment have been fxed. If two fight segments can be connected, the arrival point of one fight segment is equal to the departure point of the other fight segment, and the minimum transfer time requirement is met. In the problem of the daily fight route of the aircraft, the delay of the fight is usually related to two factors: one is due to weather reasons, trafc reasons, passenger reasons, etc.; we call it a fight delay caused by uncertainty, and the second is the next fight delay caused by the delay of the previous fight; we call it delays due to the spread of the fight. In this section, we will robustly optimize DARP based on these two aspects. We model the DARP problem. Tis problem will result in a schedule that optimizes some utility functions when there are no fight delays. In this paper, we assume a robust schedule that minimizes the cost incurred by the delay in the case of fight delays.
Based on the abovementioned remarks, we now formulate a DARP problem model. Let F(g) be the utility function of a DARP problem for any schedule g ∈ G, where G is the set of all feasible schedules. Te construction of the network is described in Section 2.1. In the following, we use the symbol v to refer to the nodes in the network graph. We defne V to be the set of all nodes in the network graph, and we assume that all fight segments are delayable. Te utility function of the network is afected by fight delays. Let F(g, v) be the utility function when segment v ∈ V is delayed. Our usual approach to dealing with this kind of uncertainty is to fnd a network with the least cost of delay in the worst case. At this point, our problem can become min g∈G max v∈V F(g, v). Defnition 1. Let G be the set of all feasible schedules let V be the set of all fight segments, and let f(g, v) be the utility function where fight v ∈ V is delayed. If there exists g * ∈ G such that min g∈G max v∈V F(g, v) � F(g * , v), we consider g * ∈ G to be an optimal robust schedule.

Network Construction.
We defne V to be the set of all legs. Each leg v ∈ V has a departure time d v and an arrival time a v . Te minimum connection time between two legs is t. For any two legs v 1 and v 2 's arrival airport is v 2 's departure airport; then, v 1 , v 2 can connect, that is, there is an arc e v 1 ,v 2 , Defne E as the set of all arcs, then a directed graph G(E, V) can be constructed.
We frst build a deterministic model that minimizes the total cost when all fights are not delayed.

Sets, Parameters, and Variables
A: A set of all matching routes in the plane and fight connection network. V: A set of all legs that need to be scheduled.
We can get the following integer linear programming (model 1): subject to In the abovementioned linear programming, the objective function (1) is to minimize the total cost when all fights are not delayed. Constraint (2) states that each fight segment can only be assigned to one aircraft. Te constraint (3) specifes that each aircraft can perform at most one route per day. Te decision variable x a is equal to 1 if the matching variable A is in the solution, 0 otherwise.

Problem Expansion.
In this subsection, a robust optimization model based on game theory is proposed to deal with the uncertainty of fight delays. Under the framework of the proposed model, two players are virtualized: one is the system maker (hereinafter referred to as player 1), and one is the attacker (hereafter also referred to as player 2). Flight delays are caused by infuences such as weather. Te scheduling problem of daily aircraft routes is formulated.

A max-min Game Model under Uncertainty.
Diferent from other robust optimization models of aircraft routes, an uncertain robust optimization model is established based on game theory. Taking player 1 and player 2 as the two participants, the constructed min-max game model is as follows: subject to where x is the decision variable, g is the decision of player 1, v is the decision of player 2, F(g, v) is the payof function, and G(x, g, v) is the constraint, G and V are the strategy sets of player 1 and player 2, respectively. In the abovementioned model, player 1 wants to minimize function F by changing g, and player 2 wants to maximize function F by changing v. From a game theory perspective, we can describe it as a noncooperative two-player zero-sum game problem. It is known from knowledge of game theory that not all twoplayer zero-sum games have a pure strategy. Nash equi- . Terefore, we adopt the min-max game model widely used in practice, as follows: subject to In the abovementioned equation, there is always a solution (g, v). In engineering technology, there are the following meanings: (a) Player 1's best strategy can handle when player 2 chooses the strategy that makes it the worst. (b) In our problem, our decision order is that player 1 chooses the strategy frst and player 2 chooses the strategy second. Player 1 needs to presuppose that player 2 chooses the worst strategy when formulating a strategy, so it is reasonable to formulate the decision problem as a min-max model. Since this is player 2's uncertainty, player 1's best strategy is to choose the worst strategy for player 2.

Modeling the Daily Airplane Route Problem as a Min-Max Game.
Te purpose of robust optimization of the daily aircraft route problem is to fnd a schedule that minimizes the cost of the system and keeps the stability of the system within a certain range. In the game model, we can think of player 2 as an attacker who wants to make the schedule made by player 1 less stable due to the uncertainty of fight delays.
In this model, one player's gain causes another player's loss, so we can simply model the model as a two-player zero-sum game. In the game model, it usually includes three elements: player set, strategy set, and payof function. Tese three elements are described in the following detail: (1) Player set: In the daily plane route problem, two players are involved, the aforementioned player 1 and player 2. Player 1 is the real maker of the schedule, whose main purpose is to fnd the schedule with the least cost. Player 2 is a virtual player with strong uncertainty, mainly afected by factors such as weather.
(2) Strategy set: Player 1's policy set is all feasible scheduling networks, i.e., g ∈ G, player 2's strategy set is all fight segments, i.e., v ∈ V. (3) Payof function: Player 1's goal is to fnd the schedule with the lowest cost while keeping the reliability of the system within a certain range. Terefore, the function value of the scheduling network is taken as the cost of player 1, and its strategy is to minimize the cost. Player 2 is deteriorating the system, and its strategy is to maximize cost.
In the daily aircraft route problem, when player 2 chooses to attack a certain fight v, the subsequent fight may also be delayed due to the spread, and this delay probability can be obtained through experience. We adopt the approach proposed by Cui et al. [10], that the probability of fight delay is related to the transit time between two fights. We assume that t 1 is the minimum connecting time specifed by the Civil Aviation Administration and m i is the probability that the connection time is between t i and t i+1 . When the connection time is greater than t k+1 , the probability of delay is m k+1 , So the probability of delay is where n is the number of fights after fight i in the aircraft's route, and q r � m i refers to the probability of the rth delay time after fight i.
Sets, parameters, and variables are as follows: So, under uncertainty, we have the following linear programming (model 2): subject to 4 Journal of Mathematics In model 2, player 2 frst delays the fight by attacking the legs, and then player 1 minimizes the total cost by adjusting the scheduling network. Te fight delay probability of player 2's attack is 1, and the delay probability of other fights is determined by their relationship with the attacked fight. When all fights are not delayed, for any segment v ∈ V, the delay probability is p v � 0, which is the same as the model in Section 2.1.
Assuming that there are multiple aircraft segments with delays and the two aircraft do not interfere with each other, we can limit each aircraft's fight path to a maximum of one aircraft that is delayed. Our model will change. So, we have the following linear programming (model 3): subject to In model 3, V 1 is a subset of V, such that for any two fight segments v 1 , v 2 ∈ V 1 , v 1 , and v 2 do not belong to any matching a at the same time. Te delay probability is 1 if v ∈ V 1 ; otherwise, the delay probability is determined by the relationship with other fights.

Solution Approach
In model 2, there are |A| decision variables, and the number of combinations of |A| increases exponentially with the increase of fights. Also, there are |V| + |K| constraints, the complexity and scale of the model are very large, so it is difcult to solve. A simple idea to solve this problem is to reduce the number of feasible schedules considered, since we can get an optimal schedule when no fight is delayed. Terefore, it is a feasible method to limit the scheduling cost to a certain proportion of the optimal scheduling. Tis makes sense, because in practical situations, a balance between cost and robustness is often pursued. For convenience, we refer to the two models above as the deterministic model (model 1) and the nondeterministic model (model 2 and model 3). Trough the abovementioned description, if we want to solve the nondeterministic model, we can frst solve the deterministic model, that is, to fnd the optimal value of the model in the absence of any fight segment delay.

Solution for Model 1.
Te branch pricing algorithm is an efcient algorithm for solving large-scale linear programming problems. Te branch-pricing algorithm is a combination of a column-generating algorithm and a branch-and-bound algorithm. Among them, the column generation algorithm can be used to solve the solution of the relaxed model. Te column generation algorithm narrows the range of candidate solutions and greatly reduces the amount of computation. After we use the column generation algorithm to obtain the candidate solution of the relaxed model, we can use the integer programming algorithm to solve it, but this often does not get the optimal solution to the original problem, so we need to use the column generation algorithm and the iteration of the branch and bound algorithm to solve it. However, in this paper, in order to save the time spent in the branch and bound process after obtaining the candidate solutions of the relaxed model; we start from the optimal solution of the integer programming and further obtain the suboptimal solution by using the variable neighborhood search algorithm. Tis greatly reduces the solution time of the model used.

Column Generation.
We frst linearly relax the aforementioned problem such that the variable x a ∈ [0, 1], and call it the restrictive main problem (RMP). In general, not all matches will appear in the optimal solution, so we only need to consider adding some matches that reduce the objective function the most to the problem model, which is called the main problem of the column generation algorithm. In fact, only a small subset of all feasible matches will be added to the model. In this way, a solution can be found quickly, then look for variables outside the model, fnd variables that can make the model better, add the model to solve again, and so on until no better variables can be added to the model.
Te way to fnd better variables is to fnd a solution that minimizes the objective function of the subproblem of the model. We set the dual variables corresponding to constraints 2 and 3 to be π v and π k , respectively. Ten, the objective function of the subproblem is minZ Note here that, c k p represents the cost of path p allocated to aircraft k , and c k represents the fxed cost of matching aircraft k contained in a, i.e., c a � c k p + c k , Terefore, we need to fnd a match that minimizes Z to add to the model. In this problem, although the number of planes is large, there may be only a few types of planes per airline. Te cost for each type of aircraft to perform the same fight segment is the same, that is, for two aircraft of the same type v 1 , v 2 ∈ V there are c k v 1 � c k v 2 . Terefore, we only need to fnd the matching k, p that minimizes Z. Tis corresponds to the shortest path problem with resource constraints, which can be solved by the labeling algorithm proposed by Si et al. [11]. We then add the resulting matches to the main problem until no match is found, such that Z < 0. In this way, we fnd all the candidate solutions corresponding to the relaxed model and then get the integer solution of the current candidate solution. But there is still a distance between the current solution and the optimal solution. [26] is an improved local search algorithm. It alternately searches using the neighborhood structure composed of diferent actions, achieving a good balance between concentration and evacuation. Te VNS algorithm mainly relies on the selection of neighborhood sets in the Shaking phase. In the Shaking phase, the algorithm randomly generates a new solution from the kth neighborhood of the solution x. Tere are 4 commonly used neighborhood actions, namely: 2-opt [23], swap-move [24], insert [25], and exchange-move. But applied to this problem, not all operations are applicable. Te algorithm designs three neighborhood actions, such as fight segment exchange, aircraft exchange, and insertion, and optimizes the solutions in sequence. Tree kinds of neighborhood action search can deeply dig out the local optimal solution, and the specifc methods are as follows:

Variable Neighborhood Search Algorithm. Variable neighborhood search (VNS)
Segment exchange: In the solution, two aircraft and segment path matches are randomly selected, and the segments in the two matched segment paths are swapped, as shown in Figure1. During this process, some exchanged solutions are infeasible, and we do not accept such infeasible solutions. Aircraft exchange: In the solution, two matches are randomly selected, and the aircraft in them are directly exchanged to form a new solution, as shown in Figure 2. Insert: In the solution, two matches are randomly selected, one or more fight segments in one of the matches are selected, and the selected fight segments are directly merged into the other match so that they are connected to form a new path, as shown in Figure 3.
Since it takes a lot of time to calculate in the branch and bound process, in order to improve the running speed, we use the variable neighborhood search algorithm in this part to get the optimal solution to the original problem. We combine the column generation algorithm and the variable neighborhood search algorithm to obtain algorithm. Steps 1-8 generate an initial solution for the column generation algorithm. Steps 9-31 improve the initial solution obtained by the column generation algorithm through domain search and fnally obtain a suboptimal solution.

Solution for Model 2.
Due to the complexity and scale of model 2, it becomes difcult to solve, but in practical applications, we cannot only consider robustness but also robustness and cost. Terefore, we need to ensure that the generated scheduling cost does not exceed a certain percentage of the suboptimal solution cost generated in Section 3.1. To solve this problem, we developed an algorithm based on variable neighborhood search. Te pseudocode for the variable neighborhood search algorithm applied to this problem is shown in algorithm. Note that F in algorithm represents the cost when there is no fight delay.
Step 1 is to generate an initial solution via algorithm.
Step 2 is to calculate the worst-case cost of the initial solution.
Step 3 is to initialize the number of iterations. Steps 4-27 are the main iterative process of the algorithm. Step 7 is the shake action. Steps 10-19 are the process of variable neighborhood descent. Steps 20-24 determine whether to update the current solution.
σ is a parameter that balances the two indicators of cost and stability. Since we describe the DARP problem as a minimization model, σ ≥ 1. When σ � 1, it means that we only choose one schedule from the optimal schedule. Te algorithm is aimed at the situation where only one defnite fight is delayed in the scheduling network. In practice, more than one fight is often delayed due to uncertainty. By making corresponding changes to the algorithm, the corresponding robust scheduling can be obtained. We can modify the single segment in algorithm to be a subset of all segments, so that we can get the solution for model 3.

Example.
Here, we give a simple example to illustrate. Te example contains 10 fight segments (for the sake of convenience, in the following description, Leg1 is written directly as L1), 2 aircraft of the same type. Tere is a certain fee for each aircraft to perform the corresponding fight segment. Also, there is a fxed fee for using the plane, which we set it as 2000. According to the China Civil Aviation Statistical Yearbook, the k1 f1 f2 f3 f4 f5 f6 k2 f7 f8 f9 f10 f11 f12 k2 f7 f8 f3 f4 f11 f12 k1 f1 f2 f9 f10 f5 f6  extra cost of one minute of delay for this type of aircraft is about 300, and the matching cost of the aircraft and fight segment is 20,000 per hour. In addition, other costs are not considered in this example. Te fight segment information is shown in Table 1. We will use the game theory approach to solve the model where the utility function is a cost. Under the condition that all fight segments are not delayed, the optimal schedule of DARP can be obtained, as shown by n 1 in Table 2. Since there is only one type of aircraft, no distinction is made between aircraft for the time being. When the delay distribution of the fight segments is unknown, we can fnd the fight segment network g * that satisfes the condition F(g * , v * ) � min g max v F(g, v), We can fnd other feasible schedules besides n 1 by setting σ � 1.05, as shown in Table 2. Table 3 lists the costs incurred by each solution when a fight segment is delayed. It can be seen from the cost table that there is no saddle point in this game, that is, there is no pure strategy Nash equilibrium. Nash equilibrium is a concept of solution in game theory. It refers to a combination of strategies that satisfy the following properties: any player who changes his strategy in this combination of strategies will not improve his own profts. We consider the more conservative case, min g∈G max v∈V � 28.17, with solutions n 4 and n 7 satisfying the condition. Hence, n 4 and n 7 are a robust schedule for this example.

Numerical Experiment
All experiments are run on an Intel(R) Core (TM) i7-1165G7 processor, 2.80 GHz, Windows 10 × 64 computer. Te code of the algorithm uses the Python programming software and adopts the CPLEX commercial processor.

Problem Scenario and Dataset Information.
Airline schedules often include multiple types of aircraft. In this setup, the airline needs to guarantee a reasonable connection time for each plane. In our experiments, the data provided by an airline in China is taken as an example to further illustrate the efectiveness of our proposed model and algorithm. For the aircraft model, the number of aircraft, the number of airports, the number of fight segments and other data included in the data, see (Table 4). As shown in Table 4, there are 7 types of aircraft in the dataset, including the "737," "73D," "73E," "73H," "73L," "73N," and "789." At the same time, more than 60 diferent airports are included in the dataset. Due to occasional aircraft maintenance, etc., the available number of aircraft in Table 4 may vary slightly. In the dataset, there are some canceled or planned canceled fights every day; in this case, we directly delete these canceled fights. Since the data of Wednesday is partially corrupted, we only use the dataset information of the remaining 6 days for the experiment. During the experiment, the dataset was divided into 6 independent datasets by date. Since our dataset contains the real scheduling information provided by airlines, it is itself a feasible solution for daily aircraft scheduling. Terefore, regardless of the airline's scheduling information, we use the basic information in the data for experimental evaluation.

Comparison of Branch Pricing Algorithm and Algorithm.
In order to highlight the gap between the enumeration method and the column generation algorithm in solving large-scale linear programming, we frst analyze the complexity of the two methods. For enumeration, it is necessary to fnd all feasible paths in the airline network. However, with the increase of fight segments, the scale of the deterministic model increases exponentially. On the contrary, for the column generation algorithm, only the paths that can reduce the objective function are found each time. Te number of paths found by the column generation algorithm is far less than the number found by the enumeration method and can quickly approach the optimal. Terefore, it is infeasible to adopt complete enumeration for a large-scale aircraft routing problem. First, we used the branch pricing algorithm and CG-VNS algorithm to test six instances, and the results are shown in Table 5. In the CG-VNS algorithm, the column generation time is shown in column CG-Time. In fact, CG-VNS and the branch pricing algorithm have the same column generation time. If we directly use CPLEX to solve all columns generated by the column generation algorithm, we can get an upper bound on the function value of the optimal integer solution. Finally, we list the gap between the CG-VNS solution and the branch pricing algorithm in the GAP column.
It can be seen from Table 5 that it is difcult to fnd largescale examples of deterministic models using the branch pricing algorithm. Te reason for the rapid increase in solution time is that the aircraft path in the connection network increases rapidly with the increase in instance size. Tis poses a great challenge to fnd high-quality solutions from a large number of decision variables. Furthermore, Table 6 reports the diference between the solution times of the VNS and CG-VNS algorithms and the solution time of the branch pricing algorithm. It can be seen from the calculation results in Table 6 that the calculation time of the VNS algorithm is lower than that of the CG-VNS algorithm. Compared with the branch pricing algorithm, the VNS algorithm is nearly 20 times faster than the standard branch pricing algorithm. However, the optimality gap of the VNS algorithm is far greater than that of the CG-VNS algorithm. Te optimality gap of the CG-VNS algorithm is less than 0.1% except, for instance, on Friday. Another advantage of our CG-VNS algorithm is that it can quickly approach the optimal solution by taking a high-quality integer solution as the initial solution of the VNS algorithm.
In a word, our algorithm achieves a good balance between time and accuracy. It is better than the branch pricing algorithm in time and the VNS algorithm in precision. In other words, we sacrifce smaller function accuracy in exchange for greater improvement in runtime. Tis is obviously acceptable for airlines.

Comparison of Model 3 and Model 1 with Flight Delays.
If a delay occurs, set the delay time to 30 minutes, and the cost per minute of delay is determined by the aircraft type. For the propagation probability of delay between two fight segments, if the transit time between the two fight segments is less than 60 minutes, the propagation probability is set to Journal of Mathematics 0.8; if the layover time is 60 minutes to 120 minutes, the propagation probability is set to 0.3; if the layover time is greater than 120 minutes, the propagation probability is set to 0.005. At the same time, we assume that only one fight on each aircraft's route is delayed due to uncertainty factors and the other fights are delayed due to propagation. In this case, we will compare the cost of the solution produced by model 3 and the solution produced by model 1 in the case of fight delays. At the same time, we also compare the costs incurred by model 1 and model 3 when the fight is not delayed.
In Figure 4, it is visually shown that the cost of daily aircraft routing problems is diferent under diferent numbers of delayed fights and using diferent models. Te abscissa in the fgure corresponds to the number of fight segments that are assumed to be delayed; that is, when the abscissa is k, it corresponds to the objective function value of model 3 corresponding to the number of fights we assume to be k. Te abscissa in Figure 4 corresponds to the number of fights assumed to be delayed. Assuming that the number of delayed fights is the green curve in Figure 4 corresponds to the worst-case cost of the scheduling generated by model 3 when the number of delayed fights is k. Obviously, when only one fight is delayed, that is, when k � 1, model 3 is equal to model 2. Te blue curve in Figure 4 corresponds to the worst-case cost of the schedule generated by model 3 when the number of delayed fights is k. In Figure 5, it shows the cost of model 3 and model 1 when the user assumes that there are k fight delays but there are actually no fight delays. As can be seen from Figure 1, the cost incurred by model 3 is higher when there are actually no fight delays. Even so, we keep the cost below 1.05x of model 1. Combining Figures 4 and 5, we conclude that model 3 is more robust when fights are delayed due to uncertain factors. Even if none of the fights are actually delayed, the schedule generated by model 3 will not cost much.

Discussion.
In conclusion, the abovementioned numerical results show that the uncertainty of fight delays will lead to an increase in airline costs. Te reason for this is that it is impossible to know in advance which fights will be delayed during aircraft routing. Terefore, airlines can only assume that some segments will be delayed in advance to avoid some losses. However, before making this assumption, we also need to estimate the number of fights with delays. If the number is too large, the cost of the airline will increase more. If the number is too small, it will cause greater losses once the delay occurs. In fact, we can judge by experience. For example, if there is a typhoon in a certain area, there is a Input: Te maximum number of iterations Maxiter; Output: Get a suboptimal solution x (1) Use the DFS algorithm to generate an initial set O and initial path; (2) while T ≠ ∅ do (3) Using the Solver to fnd the optimal solution x of RMP and dual solutions π v and π k ; (4) Change arc costs in segment network using obtained dual solution; (5) Find the matching T that minimizes Z by the labeling algorithm; end while (8) Use the solver to fnd the integer solution x of RMP; (9) t � 0; (10) while t < Maxiter do (11) k � 0; (12) while k < k max do (13) Shaking: Randomly choose a solution x′ from the k th neighbor N k (x); (14) VND: local search (15) l � 0; (16) while l < l max do (17) fnd a neighbor x″ in N l (x′); (18) if F(x″) < F(x′) then, (19) x′ � x″, l � 0; (20) else (21) l � l + 1; (22) end if (23) end while (24) if F(x′) < F(x) then, (25) x � x′, k � 0; (26) else (27) k � k + Te solution x of model 1 is obtained by algorithm 1; k � 0; (6) whilek < k max do (7) Shaking: Randomly choose a solution x′ from the kth neighbor N k (x); (8) VND: local search (9) l � 0; (10) whilel < l max do (11) fnd a neighbor x″ in N l (x′); (12) ifF(x″) < σF(x′)then (13) ifmax v∈V F(x, v) < max v∈V F(x′, v)then (14) x′ � x″, l � 0; (15) else (16) l � l + 1; (17) end if (18) end if (19) end while (20) ifmax v∈V F(x′, v) < robthen (21) rob � max v∈V F(x, v), x � x′, k � 0; (22) else (23) k � k + 1; (24) end if (25) end while (26) t � t + 1; (27) end while ALGORITHM 2: Variable neighborhood search (VNS).

Conclusion and Directions for Future Studies
Te daily aircraft routing problem is one of the biggest challenges facing the aviation industry. For the daily aircraft routing and scheduling problem, we propose three mathematical models. Te frst model is a general, nonrobust model that aims to fnd a daily minimum cost for aircraft route scheduling. We build the second model based on game theory, which considers the fight propagation delay and the delay under uncertainty. As an extension of the second le7    model, we established the third model. In the high-dimensional solution space, all three models are exponential models, so it is very difcult to efectively solve these models.
Terefore, for the general model, we also developed a VNS algorithm based on column generation. In addition, for the robust model, we also designed an improved VNS algorithm based on a column generation algorithm and proved that the proposed robust model can be efectively solved through reasonable calculations. On the one hand, a large number of real airline examples show that, compared with the exact solution generated by the branch pricing algorithm, this algorithm has a smaller gap and requires less time, so it has obvious advantages. On the other hand, compared with the nonrobust model, the robust model proposed by us has lower cost and stronger robustness.
In future research, we can consider, but not be limited to, the following directions. First, we can consider adding the impact of COVID-19 on fight delays to the model. Another interesting direction is to combine the pilot scheduling problem and the aircraft scheduling problem into a comprehensive problem because pilot scheduling [27] will also afect fight delays.

Data Availability
Te data used to support the fndings of this study are available from the corresponding author upon request.

Conflicts of Interest
Te authors declare that they have no conficts of interest.

Authors' Contributions
Bin Deng proposed the innovation and frst draft of this article; several other authors also contributed to the writing and revision of this article. Specifcally, Jingfeng Li, Junfeng Huang, and Kaiyi Tang provided experimental data for the experiment in this paper, while Weidong Li and Hao Guo did a lot of work for the revision and fnalization of this paper.