A co-evolutionary algorithm with elite archive strategy for generating diverse high-quality satellite range schedules

Satellite range scheduling, a multi-constrained combinatorial optimization problem, is crucial to guaranteeing the normal operation and application of onboard satellites. Traditional methods are dedicated to finding one optimal schedule, having ignored the problem may process multiple high-quality schedules. To provide a set of alternative schedules while maintaining the solution quality, we propose a co-evolutionary algorithm with elite archive strategy (COEAS) in this article. In COEAS, two populations are evolved to solve the original and relaxed problem in terms of schedule quality and diversity, respectively. During the evolution, the populations maintain a weak cooperation and only share the information in offspring combination phase. Further, an elite archive strategy is derived to identify and preserve potential stagnated and optimal individuals. In this strategy, the promising individuals would further participate in parent mating and offspring replacement for the dual purpose of maintaining potential optima recovery and fine-tuning the population. The experimental results show that the proposed algorithm is better than comparison algorithms in terms of efficacy (obtaining higher quality schedule), diversity (locating more optimal schedules) and flexibility (providing better alternatives).


Introduction
Satellites are equipped with various space-borne payloads to complete different military and civilian applications, such as reconnaissance and surveillance, missile warning, environmental and disaster detection, navigation and positioning, communications and broadcasting.To support these services, the satellite users request frequent satellite-ground communications with ground stations.These communications include satellite tracking, telemetry, data transmission and command betting.The satellite-ground communication is achieved through requesting visible time windows at the ground station.Then the ground station should serve the corresponding request in the given time window.The scheduling of allocating the time window to the requests is called satellite range scheduling problem [1] (SRSP).The goal of SRSP is to rationally allocate the visible time windows, satisfy the satellite user requests and resource management preference.
In SRSP, all conductions of satellite-ground communication should be subjected to the following time and resource constraints: (1) each request should be served within the visible time window; (2) a request should be served at most once; (3) an antenna supports at most one request at a time; (4) each antenna requires enough switch time before next service.Optimally schedule these multi-constrained resources into an optimal satellite range schedule has been proven to be NP-complete [2].Furthermore, the ground station antennas are oversubscribed due to the contradiction between increasing user requests and limited ground stations.Hence, finding an optimal satellite range schedule is becoming a challenging problem.
To tackle the oversubscribed characteristic of SRSP under time and resource constrains, some techniques have been developed to satisfy as much requests as possible.In sum-mary, the proposed methods can be roughly classified into the following three types: (1) E xact algorithm : Some researchers formulate the scheduling problem as a linear or nonlinear programming problem and solve it with mathematical programming approaches, such as mixed integer programming with Lagrangian-relaxation [3], branch and cut [4], exact polynomial time algorithm [5] and dynamic programming [6], etc.The exact algorithms explore the entire search space and return the same single optimal solution to the same input.As the scale of input request and resource increases, the computation burden of exact algorithms may become unacceptable due to the exponential growth of search space.(2) Local search algorithm : To reduce the search space, the local search algorithms start the search from an initial solution, iteratively search the neighbor region to improve the schedule quality and finally output a highquality schedule.For example, the greedy algorithm [7], tabu search [8] and simulated annealing [9] have been adapted to solve SRSP.Based on a conflict-resolution technique, Luo et al. [10] proposed a rescheduling strategy to rapidly improve the low-quality schedule.(3) Evolutionar y algorithm : This kind of approaches employ a population to optimize the schedule, which includes two main steps: reproduction and replacement.The overall performance enhancement of introducing objective-specified and problem-specified heuristic information into population reproduction procedure has been demonstrated in the literature [11][12][13][14][15]. Different from exact algorithms, the evolutionary algorithms avoid exploring entire search space and provide a set of solutions for decision makers to choose a final satisfying schedule.
In summary, most previous efforts concentrate on the NPcomplete nature of SRSP and seek to locate one optimal schedule, regardless of the fact that there may be multiple high-quality schedules.In SRSP, it is appealing to provide diverse optimal alternatives to decision makers for (1) quicker schedule switching under emergency events; (2) more balanced antenna load due to different choices of antennas and (3) obtaining robustness against imperfect modeling.There are several related studies of seeking diverse high-quality solutions for different problems, such as automatic itinerary planning [16], patient admission scheduling [17], traveling salesman problem [18], permutation flowshop scheduling [19], etc.However, generating multiple optimal satellite range schedules for decision makers still remains to be explored.Therefore, an effective algorithm is highly desired to solve SRSP with a set of high-quality solutions, which requires two critical issues to be considered: (1) the choice of the baseline algorithm and framework for promising search capability; (2) the technique to maintain population diversity during evolution.
Different from exact algorithm and local search algorithm, the population-based evolution algorithm solves the problem with a set of individuals, which is naturally suitable for exploring and maintaining diverse optimal solutions.However, the population tend to converge to a local optimum during the evolution, affecting the algorithm capability of finding multiple optima.The archive technique, which adapts separate archives for population convergence, diversity and potential optima preservation, has been proposed and employed to solve the problem.The two-archive algorithm [20] is the first evolutionary algorithm to promote population convergence and diversity with two collaborative archives.Li et al. [21] proposed a restricted mating selection mechanism to leverage the complementary effects of both archives, which utilize the diversity-oriented archive to explore the regions under-exploited by the convergence-oriented archive.Differently, the weak cooperation framework is adapted in [22,23], in which the two archives only interact after the offspring generation.In [24], Liu et al. developed a nichebased clearing strategy to guarantee diversity in the search space.Apart from separating the population based on convergence and diversity, the archive technique can be employed as a supporting archive during the evolution.Wang et al. [25] detects the stagnated individual with a stagnation counter and preserve these solutions in the archive.In [26], the archive individuals participate in the mutation operation to facilitate the evolution of population.
Inspired by the successful impletions of the evolutionary algorithm and archive technique, a co-evolutionary algorithm with elite archive strategy is proposed to address the above issues.The algorithm simultaneously maintains two collaborative population and an external archive: the convergence-oriented population P C , the diversity-oriented population P D and an elite archive EA.Our aim is to demonstrate that multiple high-quality schedules for SRSP can be found in a single run.The primary contributions of this work are as follows: (1) A cooperative co-evolutionary mechanism is proposed to evolve the two populations, with a relatively weak collaboration, for solving SRSP.Population P C drives the population to maintain the convergence and feasibility during evolution; P D improves the population diversity in decision space without considering the constraints.The two population only interact in the parent mating selection and offspring population combination phase.Furthermore, a learning-guided offspring generation approach is developed to generate the diversified high-quality solutions with more efficiency.
(2) An elite archive strategy is designed and implemented with the dual purpose of helping maintaining the potential optima and fine-tuning the population.EA is updated by current optimal solutions to preserve the potential optima during the evolution and protect the founded global optimal solutions.Moreover, the elite individuals are allowed to participate in the parent mating selection and population fine-tuning.To reduce the computation burden resulted from performing local search in each generation, we detect the population evolution stagnation with a stagnation counter δ.Once δ exceeds the pre-defined threshold, the replacement operation would be performed to fine-tune the population and improve the exploration ability.
(3) The experimental results validate that the above proposed strategies can improve the quality and diversity of final output schedules.Compared with the rivals, our method can obtain more diversified high-quality schedules in a single run.Different from the traditional single optima solution, the set of solutions obtained by COEAS can provide the decision makers with more alternatives to select while guaranteeing the schedule quality.The results of COEAS can be adapted as the baselines for further studies on finding diverse SRSP schedules, and our research idea can provide reference to other similar scheduling problems.
In the remainder of this article, we first describe the SRSP.
In the next section, a new cooperative co-evolutionary algorithm for the problem is presented.Experimental results and analysis is described in the subsequent section.Finally, we give our conclusions.

Problem description
Under the time and resource constraints, the satellite range scheduling problem is to determine the supported requests and the corresponding antenna allocation as well as the execution start time.In practical applications, diverse highquality satellite range schedules are desired.For example, the decision maker can select the most preferred solution between the equally good optima, instead of being limited to one option.In addition, the antenna load could be more balanced with different antenna resource allocation in each schedule.Moreover, switching to an alternative schedule would be easier and faster than re-scheduling to tackle environment uncertainty (e.g., antenna breakdown).Motivated by the application demands, we address the problem of achieving diverse high-quality satellite range schedule in a single run.
As shown in Fig. 1, the crucial point of solving SRSP is to tackle the potential conflicts in the same antenna (e.g.,

Parameter definition
We first define following parameters to describe the SRSP: where For example, the number 2 in vtw 6,2 in Fig. 1 denotes the vtw is provided by antenna a 2 , rather than the second vtw for request 6.

Mathematical model construction
Based on the above notations, the mathematical model of the SRSP could be formulated as follows.

Objective function
The objective function is to minimize the request failure rate over the planning horizon: where N R is the number of input requests and x k r ,a is a binary variable representing whether request r is supported by antenna a in its kth visible time window.

Constraints
The constraints of SRSP restrict the feasibility of the schedules.In this paper, the constraint conditions we consider are enumerated as follows: • E xecution uniqueness : Each request is supported at most once; •

Proposed algorithm
In this section, we propose a cooperative co-evolutionary algorithm with elite archive strategy (COEAS) to generate diverse high-quality solutions for SRSP.The general framework and key procedures of the algorithm would be described in turn.

General framework
COEAS evolves convergence-oriented population P C with the assistance of diversity-oriented population P D and elite archive EA.The relationships between P C , P D and EA are presented in the main loop of Fig. 2. Specifically, • P C , solves the original problem f o , is used to maintain the quality and feasibility of the solutions; • P D , solves a helper problem f h that considers less constrains, is used to assist P C in exploring the search space and jumping over the local optima; • EA, stores all elite solutions found during evolution, is used to assist P C maintain the potential optima recovery, fine-tune the population and enhance the exploration capability.
As shown in Algorithm 1 and Fig. 2, COEAS starts with randomly generated initial population P C , P D and elite archive EA as well as an initial parameter setting.In the main loop, two parent mating sets P C and P D are selected from P C ∪ EA and P D by mating selection, respectively.Then, we create new individuals from P C and P D with learning-guided offspring generation.In combination phase, the offspring population would be shared, while the fitness is evaluated by f o and f h , respectively.Afterwards, the P C , P D and EA are updated by the environmental selection and elite archive strategy.During the updating process, the population would be fine-tuned once stagnation counter δ exceeds the predefined threshold .The evolution would repeat until the maximum number of fitness evaluations is exhausted.Finally, the subset selection strategy is applied to truncate EA and obtain the final output solution set S of size μ.

Individual encoding
For an input request r , the schedule should decide: whether to support r , the corresponding time window resource and execution start time.In addition, the frequent constraint violation detect would lead to heavy computational burden.
In this paper, each individual x is encoded by the following integer array: Each request r occupies one gene of the individual, and the length of x is thus equal to the size of input requests.The gene value k represents assigning the kth VTW to request r , which is selected from 0 to K .If x r = 0, request r would not be supported.Suppose the antenna set has 3 visible time windows to support request 2, hence its variation upper bound is 3.Moreover, the gene value 3 for request 2 means vtw 3 2,4 is selected, and the resource is provided by antenna 4. The request satisfaction status and time window allocation are encoded in each individual, while the exact start time would be determined in latter decoding procedure.Furthermore, Eq. 9 guarantees each request is assigned with no more than one time window.In this way, constraint 3 and 4 can be naturally satisfied, hence the computation burden is reduced.

Mating selection
The difference between the mating selection of P C and P D lies in the union of parent population.To generate highquality offspring, we fill the mating pool P C and P D by selecting individuals from P C ∪ EA and P D , respectively.For better understanding, we formulate the input population as parent population P. Suppose solution y is the most similar solution to x, if y outperforms x in objective space, the closer distance in decision space would lead to higher possibility to select y as the mating parent.
The gene value of each request represents the serial number of assigned time window.Accordingly, the similarity between individuals is related to the number of sharing gene sites, rather than the specific gene value.The distance between solution x and y is calculated by: where x ∩ y means the intersection set of decision variables; N R is the dimension of decision variables.The smaller metric value indicates greater similarity.
The main steps for mating selection are illustrated in Algorithm 2. First, the algorithm randomly select a solution x and find the solution y ∈ P that is closest to x.Then, the parameter value θ is computed by: is satisfied, we add y instead of x to P .The above procedures are repeated until the mating pool is filled with N /2 solutions.

Offspring generation
Taking evolutionary algorithm as the baseline, the offspring generation of the proposed algorithm consists of crossover, mutation, combination and fitness evaluation.However, the randomly generated offspring often yield poor schedules in context of satellite range scheduling.
To generate diverse high-quality solutions with more efficiency, we utilize the request execution knowledge learnt from parent solution to guide the reproduction procedure.During the reproduction, the learning guidance works in three aspects: the variation probability, the variation flag and the decoding sequence.The details of the learning-guided offspring generation are given below: (1) Crossover : The multi-point crossover is adapted to randomly exchange the genes of selected sites.Given two parent solutions as shown in Fig. 4, the requests that failed to be satisfied in parent solution (e.g., request 1, 4, 5 in parent i) would be assigned a higher crossover probability of pc = min (2pc,0.95).The upper bound of pc is set to be 0.95 rather than 1, intending to avoid greedy search and premature convergence.Based on the modified crossover probability, multiple genes would be randomly selected and moved from parent j to parent i.
To further enhance the solution diversity, we designed a variation flag to detect the gene value changes.For gene r , the variation flag v f r is determined by the variation information in offspring solution and the execution information in parent solution: where x i,r = 0 denotes request r is not scheduled in parent solution i;vx i ,r is a binary variable for identifying whether gene r is variated in offspring solution i .And the default variation flag for each gene site is 3.For the variated requests, v f = 1 if were not scheduled in parent solution, or v f = 2 otherwise.
(2) Mutation : As shown in Fig. 4, the mutation operator randomly mutates multiple genes from the offspring, then determines the v f for each gene site.Unlike the crossover operator, the mutation probability is pm/N req , which has a lower occurrence probability.
(  2, the fitness of the population for f h can be easily obtained.However, the initial time window resource allocation encoded in variated operation could be infeasible in f o .Moreover, multiple individual encodings may be mapped to the same schedule by the decoding strategy.Out of this concern, we design a tri-level decoding strategy to determine the final schedule while maintaining the feasibility, quality and diversity.
During the tri-level decoding, the genes with different v f are decoded in the order of v f = 1-3.To be specific: (1) First level decoding: The v f = 1 means the request is variated in offspring solution but is not served in parent generation.Decoding these requests in first level can allocate resources to the selected failed requests with priority, which would improve the schedule quality.
(2) Second level decoding: We perform second level decoding for the requests with v f = 2.The definition of v f indicates that the resources allocated to these requests are changed, while these requests are successfully served in parent generation.
(3) Third level decoding: With regard to the requests with v f = 3, the genes corresponding to these requests remain unchanged.The decoding process is time consuming.To accelerate the decoding, we detect the conflicts between the original time window arrangement in parent schedule and current resource allocation.If the constraint is not violated, the original time window arrangement of genes with v f = 3 would be directly copied.
After deciding the decoding sequence by the v f , the request execution start time est needs to be selected within the encoded time window.Therefore, the time window allocation strategy is embedded to construct the corresponding schedule.For an input request r , the position is randomly initialized the at the top, middle and bottom of assigned time window.If the request is executable under time constraints, we determine the execution start time est, end time edt and update the resource occupied time ocupt; else we remove the request to unscheduled request set usdr.Then, we reschedule the requests in usdr based on current available time windows and ocupt.And the gene value corresponding to successfully reassigned request would be replaced with the new time window number.The reassignment would be iteratively repeated until usdr is empty or no available time window is left.Finally, the feasible schedule is obtained and the fitness can be evaluated.
To summarize, the variation operation improves the solution quality with the guidance of parent knowledge; the tri-level decoding enhances the solution diversity by decoding in diverse sequence; the time window allocation strategy guarantees the feasibility of the final schedule.
For example, in Fig. 5, suppose that the requests served on the same antenna are conflicting.Therefore, the four ground station antennas can support at most four requests.It can be observed that the initial schedule is infeasible, where request 1, 3 and request 2, 5 are conflicting.Then, we decode the solution encoding under different v f .During the i-th level decoding, we decode the genes with v f = i.The requests that conflicting with current resource occupation (requests marked in blue) would be stored and reassigned.The genes of successfully reassigned requests would be modified as the new time window number (e.g., r 2 in Fig. 5), while that of the failed requests would be assigned a 0 value (e.g., r 3 in Fig. 5).Finally, different feasible schedules are constructed by decoding under different v f .

Update mechanism of the P C
To retain randomness while keeping selection pressure towards the optima, the quality q(x) of solution x is defined by: where p_opt is the optimal solution in previous generation of P C .Then the roulette selection is used to update P C , in

Update mechanism of the P D
Different from the P C , population P D solves the easier problem and aims at generating diversified solutions.Considering less constraints indicates P D usually has better fitness than P C , which can possibly improve the performance of P C in objective space.Moreover, COEAS adapts the weak cooperation framework proposed in [22], in which the populations only interact in the combination phase.Sharing the offspring in combination operation can assist P C in crossing the infeasible regions and help P D search towards the promising region.The details of updating P D are presented in Algorithm 3. First, we split the population based on previous optimal solution: Then, population P D is updated based on the following two cases: where N k = √ 2N .Then we eliminate x r and repeat the above procedure until the size of P 1 D is decreased to N .In other words, only diverse and high-quality individuals can survive to next generation.

Update mechanism of the EA
The elite archive is in essence a memory of the promising position that the algorithm has reached in previous generations.As illustrated in Algorithm 4, the EA is updated by the combined P C instead of updated P C to avoid the effect of genetic drift.
To be specific, we compare the EA with P C to check whether the new elite solutions are generated.A new solution Moreover, we define the minimal archive size μ to help the algorithm explore the search space with different behavior.At the early iterations, the elite solutions are directly added to the archive until the archive size reaches to μ.These widely distributed historical optimal solutions h_opt enables the algorithm to better explore the search space.Then we gradually eliminate the worst h_opt through truncation.In the late iterations, the archive is filled with multiple local optima.Restarting the search from these individuals can help the algorithm locate better optimal solutions.

Replacement operation
The population individuals tend to converge to the earlyfound local optima during the evolution, which eventually leads to the premature convergence and diversity lost.Out of this concern, we detect and count the population stagnation in previous update procedure.Once the stagnation counter δ exceeds the predefined threshold , we replace the worst/crowded solution in P C /P D with neighbor solution of EA to fine-tune the population.
The procedures of replacement operation are described in Algorithm 5. First, the algorithm randomly selects N elite solutions from EA, and finds the solution x c with worst fitness in P C , the most crowded solution x d in P D .Next, the neighbor region of the selected solutions is exploited through mutation operation.The neighbor solution with best fitness and maximum distance to P D is denoted by x cn and x dn , respectively.In the following, we compare the fitness of x cn and x c .If the fitness of x cn is better than x c or a randomly generated number is less than a threshold ξ , we reset δ and replace the x c and x d with x cn and x dn , respectively.Otherwise, the above procedures would be repeated for five times until the replacement is conducted.The threshold ξ is com- Moreover, the ability of revisiting parts of the search space can yield a better balance between seeking greedily for new solutions and seeking for high-quality solutions locally [27].In the proposed algorithm, the elite archive individuals are allowed to participate in the mating selection of P C for the purpose of maintaining the potential optima recovery.

Experimental instances
Based on actual engineering, the generated experimental instances are involved with both LEO, MEO, GEO satellites and globally distributed ground station.Part of the ground stations is equipped with multiple antennas to support the requests, which can increase the possibility of finding multiple optimal schedules.The size of input requests ranges from 150 to 350 and the scheduling horizon is set to be 24 h.

Contrast algorithms
Three algorithms are selected for performance comparison: (a) CCMO [22]: a coevolutionary constrained multiobjective optimization algorithm, which first proposed the weak-cooperation framework; (b) GA-PE [28]: a genetic algorithm with population perturbation and elimination strategy for SRSP; (c) GA [29]: a classic evolutionary algorithm, which is adapted as the performance baseline.
For fair comparison, each algorithm adapts the same default parameter settings as shown in termination criterion is the maximum fitness evaluations (MaxFEs).Moreover, the elite archive is incorporated to the comparison algorithms for the purpose of preserving the historical optimal solutions.Limited to the search ability, the final archive size for each algorithm in different runs varies.Hence, we modify the output of each algorithm as μ elite solutions from E A, instead of the final population.And the representative μ elite solutions are selected by the truncation operation in elite archive strategy.

Performance indicators
Three evaluation indicators are adapted to measure the quality and diversity of obtained solutions: (1) F β Measure F β Measure, derives from information retrieval and pattern recognition [30], is a comprehensive indicator to evaluate the solution quality.For a ground-truth solution set G, the F β measure of solution set S is calculated as where β 2 is a parameter to control the effect of precision value pr ec and the recall value rec on the metric value.In this paper, the β 2 is set to 0.3 as advised in [31].The pr ec shows the fraction of optimal solutions in S: where TP and FP are the number of optimal and non-optimal solutions in S.And the rec denotes the fraction of optimal solutions the algorithm found in G: where F N are the number of optimal solutions the algorithm failed to locate in G.
Particularly, the F β score equals to 1/0 when all/none of the solutions in S are optimal.
(2) Diver sit y I ndicator The F β scores for the algorithms, which fail to locate any problem optima, are 0. To further distinguish the algorithm performance, the diversity indicator (DI) is adapted to measure the average maximum similarity between S and G in decision space: The DI describes the convergence degree of the obtained solutions in decision space.And the greater metric value indicates the better performance.
The above two indicators give a comprehensive comparison on finding multiple optimal solutions, and more detailed descriptions can be found in [18,31].During the performance evaluation, the final optimal solution set that all the algorithms have obtained for each instance is adapted as an approximation to the ground-truth solution set G.
(3) Request Failure Rate The request failure rate (RFR) evaluates the output solution set S in terms of satisfying the input requests: The RFR shows the convergence degree of obtained solutions in objective space (i.e., the capability to optimize the SRSP).

Exploring the proposed method
To analyze the significance of the proposed elite archive strategy, learning-guided offspring generation and mating selection in COEAS, we compare the COEAS with its five variants: COEAS with fixed elite archive size (COEAS-f); COEAS that adapts global optima as the population update reference point (COEAS-g); COEAS that randomly generates the offspring (COEAS-r); COEAS without replacement operation (COEAS-nr) and COEAS without parent mating selection (COEAS-nm).Table 2 presents the mean metric value comparisons from 20 runs for each algorithm.For each instance, the best and second-best result is highlighted in bold and italicized, respectively.The Wilcoxon rank-sum test is applied to further investigate the effect at significant level α = 0.05, where symbols '+', '−', and '≈' indicates that the algorithm performance is significantly better than, significantly worse than, or similar to that of COEAS.
As seen in the table, the proposed strategies can effectively enhance the overall performance of COEAS.Specifically, the The proposed method is compared with its five variations COEAS-f and COEAS-g achieve most of the second-best performance.Without learning-guided offspring generation, the COEAS-r fails to locate the optima in given iterations.Due to the limitation of multi-constrained and combinatorial characteristics of SRSP, the initial randomly generated solutions could hardly maintain quality and feasibility, simultaneously.Without replacement operation, COEAS-nr obtains the second-best performance on instance 275 and 300, and performs badly with respect to the RFR value.Consequently, by exploiting the neighbor region of elite archive candidates, the replacement operation can effectively improve the population quality.COEAS-nm cancels the mating selection of parent solutions and directly reproduces new solutions from P C and P D , respectively.In terms of F β and DI, the COEASnm shows worse performance, which indicates the worse ability to explore the search space.This result is due to the potential optima recovered from historical optima in E A, and the competition between the similar solutions.
To present a further investigation of the performance concerning different indicators, the mean metric value for all algorithm variations on instance 225 is shown in Fig. 6.At early stage of the evolution, COEA-f and COEAS-g converge faster than COEAS.As the evolution progresses, COEAS outperforms the variations, which shows that the assistance of P D and E A can help P C better explore the search space and facilitate the population evolution.As for COEAS-nm, it is outperformed by all the peer variations except COEAS-r.Its failure reflects that the interactions between P C and E A in parent mating can effectively avert being trapped in local optima.

Comparing with the related algorithm
Then, we compare COEAS with recently proposed methods including coevolutionary algorithm that adapts weak cooperation framework (CCMO) [22], population perturbation and elimination strategy based genetic algorithm (GA-PE) [28] and classical genetic algorithm (GA) [29].As analyzed in previous section, the randomly generated offspring in original evolutionary algorithm has difficulty to solve SRSP with efficiency.To make a meaningful comparison, the proposed offspring generation approach is implemented in all algorithms on each instance for efficiency enhancement.Table 3 provides the comparison results of each algorithm.The original and modified version of the algorithm is denoted by 'random' and '/', respectively.The best result for each instance and the better result within each pair of algorithms is highlighted with bold and italicized, respectively.It can be observed that COEAS processes the superior performance over compared algorithms in terms of F β , DI and RFR.Specifically, (1) The F β score for the modified version of each algorithm significantly outperforms that of the original algorithm, which validates that better solving efficiency can be achieved with the guidance of heuristic information learnt from parent schedule.For each instance.COEAS reaches the highest F β value and provides most of the optimal solutions.CCMO achieves the second-best results in most instances.The improved GA-PE and GA can find a limited number of optimal schedules, but still perform poorly on instances with larger scale.
(2) Higher DI indicates the better solution diversity.COEAS can maintain the best diversity among the algorithms.GA-PE and GA show similar performance on providing diverse solutions, while CCMO presents better performance by providing more distinct optima.
(3) Request failure rate reflects the ability of the algorithm to tackle SRSP.We present a quick view of RFR score by plot-Fig.7 The average RFR values of the solutions obtained by COEAS, CCMO, GA-PE and GA ting the average metric value for different instances in Fig. 7.The corresponding original version is marked with 'r'.It can be observed that the final output solutions of COEAS have a good convergence to the optimum of each instance.And all the algorithms are encountered with difficulty to satisfy more requests for the instances with a large size.
To present an intuitive algorithm performance comparison concerning F β , DI and RFR, we adapt the performance indicator p(Alg i ).For algorithm Alg i , p(Alg i ) reflects the number of algorithms that are significantly better than Alg i .The lower indicator value means better algorithm capability.More detailed description of Alg i can be found in [32].It can be observed that COEAS and CCMO turn to have the best performance.The basic GA-PE outperforms basic GA, while the overall performances are closed with the introduction of heuristic information.
Figure 8 compares the quality, diversity and convergence of elite archive solutions on instance 225.It can be seen

Influence of replacement threshold
To investigate the effect of replacement threshold, we conduct the parameter sensitivity verification experiments in this section.Six groups of experiments, corresponding to Particularly, the value 0/∞ means replacement operation in each/no generation, respectively.For the sake of brevity, we present the mean performance score of different in terms of F β , DI and RFR on different instances in Fig. 9. Compared with = ∞, the proposed replacement operation can help the algorithm to maintain good overall performance in terms of both solution quality and diversity on different, especially large-scale, instances.In regard of = 0, the frequent local search can handle the instances with small size, but have difficulty to tackle the large-scale instances.at other parameter levels show similar performance on most instances.
In Fig. 10, we further compare the algorithm execution time, which rises dramatically along with the increasing input request size.In the replacement operation, the local search would be repeated for five times until a desired solution is found.= 0 allocates most computational resource to find the neighbor solution, hence consumes the least execution time on each instance.In this paper, the is finally assigned to 2 for the purpose of fine-tuning population quality and avoiding unnecessary exploitation.

Effect of obtaining multiple solutions for SRSP
In this section, the effect of providing multiple high-quality solutions for decision makers are investigated.
To describe the preference of decision makers for selecting the final schedule, the following two indicators are adapted: (1) Load Balance Rate: It evaluates the working load balance rate of each ground station antenna.
where I the set of emergency input requests; N i means the number of original requests that are affected by request i.The higher schedule stability rate means the ground station resource can tackle the emergencies with faster response.
During the experiment, the single best solution from final population P C is utilized as the baseline.For SSR analysis, we randomly input emergency requests with 10% of the instance size.Figure 11 provides the comparison results.It can be seen that COEAS possesses the superior performance over traditional method.The diversified high-quality schedules can provide the decision makers with more alternatives to select based on their practical concerns.

Conclusion
This paper develops a co-evolutionary algorithm with elite archive strategy for tackling SRSPs, which can provide a set of diverse high-quality schedules in single Two techniques are effectively employed to improve the perof the algorithm: (1) cooperative co-evolutionary mechanism; (2) elite archive strategy.The cooperative coevolutionary mechanism drives two populations explore the problem space during evolution with a weak collaboration.The elite archive strategy helps identify and preserve potential optima, so as to avoid the effect of genetic drift.Moreover, the archive individuals participate in parent mating for enabling the algorithm to revisit the historical promising region.The replacement operation adaptively fine-tunes the population to alleviate the premature convergence and avert unnecessary local search.To further enhance the search ability, the offspring generation approach reinitializes and decodes the new schedules with the guidance of request satisfaction information in parent schedule, so as to improve the quality and diversity of the new schedules.
Based on these techniques, the COEAS shows a promising performance both in reaching better objective value and locating more optimal schedules per run.The experimental results also the flexibility of COEAS for providing more high-quality alternative schedules.
For future works, we will extend the algorithm to tackle more complicated instances, such as instances with larger request scales or considering more objectives.In addition, some effective techniques can be further explored and utilized to solve the problem, such as niching techniques.cate if changes were made.The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material.If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.To view a copy of this licence, visit http://creativecomm ons.org/licenses/by/4.0/.

Fig. 1 A
Fig. 1 A SRSP of six requests in a discrete time period

Fig. 2
Fig.2The flowchart of the proposed algorithm Fig 3 gives a simple example of individual encoding.

Fig. 3
Fig. 3 Simple example for individual encoding

)
Combination and Fitness Evaluation : First, the current parent population P C and P D is combined with offspring populations O C ∪ O D .Then, the fitness of O C ∪ O D in updated P C and P D is evaluated by f o and f h , respectively.Based on Equation

1 vtw 1 Fig. 5
Fig.5 The decoding of a solution under different variation flags

Fig. 6
Fig. 6 Mean metric value for different algorithm variations during the evolution

Fig. 9 Fig. 10
Fig. 9 Mean performance score of different on different instances

Fig. 11
Fig. 11 Performance comparisons between obtaining multiple high-quality solutions and finding single optima r , ...N R } denotes the set of N R requests.And the vtw is provided by antenna a.
If the new elite solutions are not detected, we consider the population evolution is stagnated, and add 1 to the stagnation counter δ.If the new Algorithm 4 Update Mechanism of the E A Elite archive EA, Combined P C , Stagnation counter δ, Minimal archive size μ, Previous optima set p_opt Ensure: Updated archive EA 1: [s f , c_opt] ⇐ Stagnation_Detection(EA, P C ) Require: Elite archive E A, P C , P D ,Stagnation counter δ Ensure: Updated P C , P D , δ 1: if δ > then 2: [x c , x d ] ⇐ Identification(P C , P D ) C , P D ] ⇐ Replacement(P C , P D , x c , x d , x cn , x dn ) Require:

Table 1
The default algorithm parameter settings

Table 2
Average metric score in terms of F β , DI and RFR on each instance over 20 runs

Table 3
Average metric score in terms of F β , DI and RFR on each instance over 20 runs.Mean metric value for different related algorithms during the evolution that COEAS can identify and preserve the optimal schedules with compared algorithms, COEAS can jump out of the local optimum easily with the assistance of P D and E A.