Incremental learning enhanced integrated proxy-assisted evolutionary algorithm

Abstract

The invention discloses an incremental learning enhanced integrated agent assisted evolutionary algorithm, which comprises the following steps of: setting parameters of a self-adaptive rolling time domain strategy according to information of a ship to be allocated, and dividing a long-term berth quay crane collaborative allocation problem (LBACAP) into continuous sub-LBACAPs; solving the sub LBACAP by using an integrated agent assisted evolutionary algorithm; the arrival time and departure time are predicted according to the arrival sequence of the ships, dynamic adjustment and processing of the continuous arrival ships are achieved, time windows are divided, and it is ensured that the algorithm can adapt to the continuous arrival field of the ships by dynamically adjusting the time windows; three proxy models of Kriging, RBFN and SVR are utilized, so that the calculation cost is remarkably reduced, and meanwhile, the strong search capability is kept; historical decision information is used for guiding the optimization process of a subsequent time window, and the optimization effect of the algorithm is gradually improved; through combination of a rolling window and incremental learning, an optimal berth and quay crane distribution scheme can be found in a global range.

Description

Technical Field

[0001] The invention belongs to the technical field of port berth allocation, and in particular is an integrated agent-assisted evolutionary algorithm enhanced by incremental learning. Background Art

[0002] In global trade, maritime transport occupies an important position. The operating efficiency of container ports affects the economic benefits and timeliness of international logistics. In port operations, the Berth Allocation and Quay Crane Assignment Problem (BACAP) is a key task. Its optimization goal is to reduce ship waiting time and improve operating efficiency by reasonably arranging the berth location and quay crane resources of ships, thereby reducing operating costs. However, due to the uncertainty of ship arrival time, limited resources and interdependence between ships, the traditional method of solving BACAP faces many challenges in practical applications. In actual port operations, ships do not arrive at the port in a fixed time period, but arrive in a continuous manner. This feature is the key problem to be solved in this patent, namely the Lifelong Berth Allocation and Quay Crane Assignment Problem (LBACAP), which is far more dynamic and complex than traditional BACAP.

[0003] Most existing studies focus on BACAP optimization in fixed time periods, and evolutionary algorithms (EAs) are widely used as core solution technologies. However, these traditional methods face significant limitations when dealing with LBACAP. First, most studies are based on the assumption of fixed-scale scenarios, that is, the default number of ships is fixed, and the dynamic nature of the continuous arrival of ships in port operations is not fully considered, resulting in insufficient performance when dealing with long-term scenarios. Secondly, many methods rely on genotype-encoded decoding strategies. Although this strategy can theoretically obtain a better solution, it requires a complex calculation process, and the computational cost increases significantly with the growth of the problem scale, thereby limiting the practicality of the algorithm in dynamic and large-scale scenarios. In addition, some studies use the method of decomposing the problem into multiple independent small-scale sub-problems to reduce the difficulty of solving, but this method usually ignores the interrelationships between ship allocation schemes in different time windows, resulting in the failure to fully capture the complex dependencies across time windows in LBACAP, thereby affecting the global optimization effect. Therefore, this patent proposes an incremental learning-enhanced integrated agent-assisted evolutionary algorithm framework for LBACAP in response to the above-mentioned limitations. This patent uses an integrated agent-assisted evolutionary algorithm framework enhanced by incremental learning to handle the fluctuation of ship arrival time and the limited resources, effectively improving the operational efficiency of ports (especially busy international trade ports), and playing an important role in the global supply chain. This algorithm framework is not only applicable to ports, but can also be extended to other scenarios that require dynamic allocation and scheduling of resources, such as air freight, rail freight, etc., and has practical value across industries. Summary of the invention

[0004] Aiming at the Lifelong Berth Allocation and Quay Crane Assignment Problem (LBACAP), the present invention proposes an Incremental Learning-Enhanced Ensemble Surrogate-Assisted Evolutionary Algorithm (IL-ESAEA) framework, which provides a new technical solution for port scheduling optimization in continuous scenarios.

[0005] In order to solve the above problems, the present invention provides a technical solution:

[0006] An incremental learning enhanced integrated agent-assisted evolutionary algorithm, comprising the following steps:

[0007] S1. According to the information of the ship to be allocated, set the parameters of the adaptive rolling time domain strategy and divide the LBACAP into continuous sub-LBACAPs;

[0008] S2, solving the sub-LBACAP using an integrated agent-assisted evolutionary algorithm;

[0009] S3. Estimated arrival and departure times according to the order of ships arriving at the port, and dynamic adjustment and processing of ships that continue to arrive at the port;

[0010] S4, divide the time window, and implement the incremental learning mechanism in the second and subsequent individual initialization processes of each time window;

[0011] S5. Repeat the rolling window optimization and population update process until all ships are assigned;

[0012] S6. Integrate the optimization results of all time windows and output the global optimal berth and quay crane allocation plan.

[0013] Preferably, the parameters of the adaptive rolling time domain strategy in S1 are: the number of ships W in the time window and the step size η of the rolling window.

[0014] Preferably, the step of solving the sub-LBACAP in the h-th time window in S2 is:

[0015] S21, randomly generate N according to the berthing order of the ships to be berthed init Initial individuals, each of which contains different berthing order information. By decoding each individual, its corresponding berth allocation plan and quay crane scheduling plan are obtained, and its true evaluation value is calculated;

[0016] S22, the initial N init Individuals are added to the training set S of the proxy model, where the training set is an empty set at the beginning, and then the training set is used to train three proxy models respectively;

[0017] S23, select the best NP individuals from S according to the true evaluation value to form the initial parent population. Through selection, crossover and mutation operations, the parent population generates the child population;

[0018] S24, using the trained proxy model to predict the performance of the offspring individuals;

[0019] S25. According to the prediction results, select q individuals from the offspring to form a supplementary individual set Q, decode the individuals in Q, obtain their true evaluation values, and update the current optimal solution;

[0020] S26, merge Q and S to obtain a new training set, and use the updated training set to update the three proxy models;

[0021] S27, repeat S21-S26 until the stop criterion is met;

[0022] S28. Select the solution obtained by the model with the highest accuracy among the three proxy models as the final solution.

[0023] Preferably, the specific implementation steps of S3 are:

[0024] S31. According to the order of ships arriving at the port, the first time window ranges from v 1 Estimated time of arrival at port is v W The estimated departure time of the port is calculated and the sub-LBACAP solution in the first time window is completed;

[0025] S32, the first n ships (i.e. v 1 to v η ) are added to the final dequeue, and the remaining (W-η) ships are added to the next time window for optimization together with the newly arrived ships;

[0026] By adding the first n ships waiting to be dispatched, the number of ships in the time window becomes v η+1 to v W+η , the range of the time window is updated according to formulas (1) and (2):

[0027] tb h = min a Vh , (1),

[0028] te h = max tl Vh , (2),

[0029] Among them, a Vh represents the set of estimated arrival times of ships in the hth time window, tl Vh represents the set of estimated departure times of ships in the hth time window, tb h and h Respectively represent the start time and end time of the hth time window;

[0030] S33. Scroll the time window forward to achieve dynamic adjustment and processing of the ships that continue to arrive at the port.

[0031] Preferably, the specific implementation steps of S4 are:

[0032] S41, using the solution x* obtained in the previous time window h Guide the generation of the initial individuals in the current time window;

[0033] S42, x* h The overlapping part in the time window is taken as the reference sequence and randomly combined with the order of the newly arrived ships to be scheduled.

[0034] The beneficial effects of the present invention are:

[0035] It can solve long-term LBACAP, is applicable to dynamic and complex situations in actual port operations, and has high practical value;

[0036] Through the adaptive rolling horizon strategy and incremental learning mechanism, the algorithm can quickly respond to newly arrived ships, reduce computing costs, and improve optimization efficiency;

[0037] The integrated agent model can effectively handle the differences in ship information distribution in different time windows and improve the robustness and adaptability of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] For ease of explanation, the present invention is described in detail with reference to the following specific implementations and the accompanying drawings.

[0039] Figure 1 is a flow chart of the incremental learning enhanced integrated agent-assisted evolutionary algorithm framework of the present invention;

[0040] Figure 2 2 is a diagram of experimental results of an embodiment of the present invention. DETAILED DESCRIPTION

[0041] like Figure 1-2 As shown, the specific implementation adopts the following technical solutions:

[0042] Embodiment 1:

[0043] An incremental learning enhanced integrated agent-assisted evolutionary algorithm, comprising the following steps:

[0044] S1. According to the information of the ship to be allocated, set the parameters of the adaptive rolling time domain strategy and divide the LBACAP into continuous sub-LBACAPs;

[0045] S2, solving the sub-LBACAP using an integrated agent-assisted evolutionary algorithm;

[0046] S3. Estimated arrival and departure times according to the order of ships arriving at the port, and dynamic adjustment and processing of ships that continue to arrive at the port;

[0047] S4, divide the time window, and implement the incremental learning mechanism in the second and subsequent individual initialization processes of each time window;

[0048] S5. Repeat the rolling window optimization and population update process until all ships are assigned;

[0049] S6. Integrate the optimization results of all time windows and output the global optimal berth and quay crane allocation plan.

[0050] Through the organic combination of adaptive rolling time window strategy, integrated agent-assisted evolutionary algorithm and incremental learning mechanism, an efficient and practical optimization method is provided for the long-term berth and quay crane allocation problem. This algorithm has important theoretical value and broad practical application prospects in the field of port scheduling, which helps to improve the efficiency of port resource allocation, reduce operating costs, and provide a new solution for the intelligent management of modern ports.

[0051] Embodiment 2:

[0052] In order to verify the effectiveness of the present invention, this example applies the algorithm proposed by the present invention to this example to give a final explanation. In the example, the length of the shoreline is 3000 meters, there are 30 quay cranes on the shore, the working efficiency of the quay cranes is 25 boxes / hour, and berths and quay cranes need to be allocated for 100 ships arriving at the port within the next 400 hours;

[0053] The input parameter symbols and settings of the ship are shown in the following table:

[0054]

[0055]

[0056] According to this setting, 10 different data sets are generated. In this embodiment, the optimization objective function established is as follows, with the purpose of minimizing the waiting time for the ship to berth and the offset distance of the target berth:

[0057] min∑ i＝1 α(ts i -a i )+β|p i -b i |, (3)

[0058] Where α = 10, β = 1, a i -ts i Indicates the waiting time for parking, b i -p i Indicates the target berth offset distance;

[0059] First, according to the information of the ships to be assigned in the example, the parameters of the ARH strategy are set: the number of ships W in the time window and the step size η of the rolling window, and the LBACAP is divided into continuous sub-LBACAPs;

[0060] Subsequently, the ensemble agent-assisted evolutionary algorithm was used to solve the sub-LBACAP problem instance. N init Initial individuals, each of which contains different berthing order information. By decoding each individual, its corresponding berth allocation plan and quay crane scheduling plan are obtained, and its true evaluation value is calculated.init Individuals are added to the training set S of the proxy model, where the training set is an empty set at the beginning. Then, the three proxy models are trained respectively with the training set. Next, the best NP individuals are selected from S according to the true evaluation value to form the initial parent population. Through selection, crossover and mutation operations, the parent population generates the child population. Subsequently, the trained proxy model is used to predict the performance of the child individuals. According to the prediction results, q individuals are selected from the offspring to form a supplementary individual set Q. The individuals in Q are decoded to obtain their true evaluation values, and the current optimal solution is updated. Q is merged with S to obtain a new training set, and the updated training set is used to update the three proxy models. This process is repeated until the stopping criterion is met, and the solution obtained by the model with the highest accuracy among the three proxy models is selected as the final solution.

[0061] Next, the time window is rolled over. After the sub-LBACAP solution in the current time window is completed, the first n ships are added to the final solution queue, and the remaining (W-n) ships are added to the next time window together with the newly arrived ships for optimization. The range of the time window is updated according to formulas (1) and (2):

[0062] tb h = min a Vh , (1),

[0063] te h = max tl Vh , (2),

[0064] Among them, a Vh represents the set of estimated arrival times of ships in the hth time window, tl Vh represents the set of estimated departure times of ships in the hth time window, tb h and h They represent the start time and end time of the hth time window respectively. Then, the time window rolls forward to achieve dynamic adjustment and processing of the ships that continue to arrive at the port;

[0065] Then, incremental learning and population update are performed. Since the overlapping parts set between adjacent time windows will cause the ships in the overlapping parts to need to be repeatedly scheduled and calculated, the incremental learning mechanism is implemented in the second and subsequent individual initialization processes of each time window, using the solution x obtained in the previous time window. * h To guide the generation of the initial individual in the current time window, x * h The overlapping part in the time window is taken as the reference sequence and randomly combined with the order of the newly arrived ships to be scheduled, such as Figure 1As shown in Figure 1, in part A, when the scheduling of ships 1 to 10 in the time window h is completed, the final berth order is determined to be {1,2,3,4,5,6,7,8,9,10}, among which the berth order of ships 1 to 5 {1,2,3,4,5} is added to the final solution sequence, while ships 6 to 10 carry incremental information {6,7,8,9,10} and participate in a new round of scheduling optimization together with the newly arrived ships 11 to 15. Part C shows the specific use of incremental information. When the initial individual is generated in the time window h+1, the incremental information {6,7,8,9,10} is regarded as the benchmark sequence, and the total number of benchmark sequences is Add one insertion point, and ships 11 to 15 will randomly select insertion points in turn. For each inserted ship, the benchmark sequence will be updated accordingly. All newly arrived ships will be inserted to generate an individual. Since the incremental learning mechanism accelerates the convergence speed of each time window, an adaptive iteration stopping criterion is designed to save unnecessary computing resources. Specifically, when no better individual is generated in x consecutive iterations, the population will be reinitialized and the iteration will be restarted to prevent the algorithm from falling into a local optimal solution. Based on the above principle, if no better individual is generated in y consecutive iterations (y>x), the iteration process will be stopped;

[0066] Finally, the iterative solution is performed, and the rolling window optimization and population update process is repeated until all ships are assigned and the results are output. The optimization results of all time windows are combined to output the global optimal berth and quay crane allocation plan;

[0067] The algorithm input parameters are set as follows: population size pop_num = 100, crossover rate crossover_rate = 0.8, mutation rate muation_rate = 0.4, number of clusters K = 4, diversity enhancement frequency G = 50, time window size W = 10, rolling step η = 5, maximum number of evaluations per time window maxEval = 200, thresholds of iterative stopping criteria x = 50, y = 70, initial number of individuals N init =100, the number of individuals added in each iteration is q=1;

[0068] The final experimental results (optimal solution) are represented by a diagram, in which the horizontal axis represents the coastline, the vertical axis represents the time, and each solid rectangular block represents the berthing situation of a ship (the horizontal axis in the lower left corner of the rectangle represents the berthing position, the vertical axis represents the berthing time, the width of the rectangle represents the length of the ship, and the height represents the working time). The three numbers in the rectangle represent the ship number, the ship offset distance and the waiting time, respectively. The dotted box represents the safe distance of the ship (10% of the length of the ship), and the black dot represents the estimated arrival time and target berth of each ship. The arrow can reflect the size of the ship offset target value.

[0069] In order to verify the performance of the algorithm framework, the algorithm HMA is combined with the IL-ESAEA framework and compared with SAGA (agent-assisted genetic algorithm), HMA (memetic algorithm of heuristic decoding method) and ALNS (adaptive large domain search algorithm) in the above examples. Each algorithm is run 20 times on each example to obtain the target value of its solution;

[0070] Depend on Figure 2 It can be seen that the algorithm IL-ESAEA proposed in the patent of this invention has achieved the best results on all data sets, and the calculation speed of IL-ESAEA is also significantly higher than that of HMA and ALNS. SAGA benefits from the optimization of algorithm efficiency brought by the agent model, which greatly reduces the running time without reducing the algorithm performance. However, the performance of SAGA is far behind that of IL-ESAEA. Therefore, IL-ESAEA has strong competitiveness in both performance and efficiency, and can provide excellent solutions in a short time. It can be seen that the incremental learning enhanced integrated agent-assisted evolutionary algorithm proposed in this invention can significantly improve the response speed and flexibility of port operations in a diverse and complex port operation environment, and reduce the losses caused by calculation delays. Its excellent solution capability helps to cope with the ever-changing shipping needs and ensure the continuous stability and efficient operation of port operations.

[0071] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An integrated agent-assisted evolutionary algorithm enhanced by incremental learning, characterized in that: The steps include: S1. According to the information of the ship to be allocated, set the parameters of the adaptive rolling time domain strategy and divide the LBACAP into continuous sub-LBACAPs; S2, solving the sub-LBACAP using an integrated agent-assisted evolutionary algorithm; S3. Estimated arrival and departure times according to the order of ships arriving at the port, and dynamic adjustment and processing of ships that continue to arrive at the port; S4, divide the time window, and implement the incremental learning mechanism in the second and subsequent individual initialization processes of each time window; S5. Repeat the rolling window optimization and population update process until all ships are assigned; S6. Integrate the optimization results of all time windows and output the global optimal berth and quay crane allocation plan.

2. The incremental learning enhanced integrated agent-assisted evolutionary algorithm according to claim 1, characterized in that: The parameters of the adaptive rolling time domain strategy in S1 are: the number of ships W in the time window and the step size η of the rolling window.

3. The incremental learning enhanced integrated agent-assisted evolutionary algorithm according to claim 1, characterized in that: The steps for solving the sub-LBACAP in the h-th time window in S2 are: S21, randomly generate N according to the berthing order of the ships to be berthed init Initial individuals are decoded to obtain the corresponding berth allocation plan and quay crane scheduling plan, and their true evaluation values ​​are calculated; S22, the initial N init Individuals are added to the training set S of the proxy model, where the training set is an empty set at the beginning, and then the training set is used to train three proxy models respectively; S23, select the best NP individuals from S according to the true evaluation value to form the initial parent population. Through selection, crossover and mutation operations, the parent population generates the child population; S24, using the trained proxy model to predict the performance of the offspring individuals; S25. According to the prediction results, select q individuals from the offspring to form a supplementary individual set Q, decode the individuals in Q, obtain their true evaluation values, and update the current optimal solution; S26, merge Q and S to obtain a new training set, and use the updated training set to update the three proxy models; S27, repeat S21-S26 until the stop criterion is met; S28. Select the solution obtained by the model with the highest accuracy among the three proxy models as the final solution.

4. The incremental learning enhanced integrated agent-assisted evolutionary algorithm according to claim 1, characterized in that: The specific implementation steps of S3 are: S31, according to the order of ship arrival, the first time window ranges from the estimated arrival time of v1 to v W The estimated departure time of the port is calculated and the sub-LBACAP solution in the first time window is completed; By adding the first n ships waiting to be dispatched, the number of ships in the time window becomes v η+1 to v W+η , the range of the time window is updated according to formulas (1) and (2): tb h = my a Vh , (1), you h = max tl Vh , (2), Among them, a Vh represents the set of estimated arrival times of ships in the hth time window, tl Vh represents the set of estimated departure times of ships in the hth time window, tb h and h Respectively represent the start time and end time of the hth time window; S32, adding the first n ships to the final dequeue, and adding the remaining (W-n) ships and the newly arrived ships to the next time window for optimization; S33. Scroll the time window forward to achieve dynamic adjustment and processing of the ships that continue to arrive at the port.

5. The incremental learning enhanced integrated agent-assisted evolutionary algorithm according to claim 1, characterized in that: The specific implementation steps of S4 are: S41, using the solution x* obtained in the previous time window h Guide the generation of the initial individuals in the current time window; S42, x* h The overlapping part in the time window is taken as the reference sequence and randomly combined with the order of the newly arrived ships to be scheduled.

Images