Multi-domain oriented joint strategy selection and budget allocation method and system

Abstract

The application discloses a multi-domain-oriented joint strategy selection and budget allocation method and system, relates to the technical field of artificial intelligence and machine learning, and comprises the following steps: acquiring an unlabeled sample pool divided into a plurality of heterogeneous domains in advance and historical learning gains of each heterogeneous domain; performing feedback-driven domain budget allocation calculation based on the historical learning gains of each heterogeneous domain to obtain query sample quotas of the domains; performing strategy selection through preset Thompson sampling based on the query sample quotas of the domains to obtain an optimal strategy; and screening the unlabeled sample pool of the plurality of heterogeneous domains based on the optimal strategy to obtain candidate samples; and calculating a budget-aware composite reward of the candidate samples based on a preset gating coefficient, so that higher classification accuracy and lower labeling cost are achieved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and machine learning technology, specifically to a method and system for joint strategy selection and budget allocation across multiple domains. Background Technology

[0002] In real-world applications, training deep neural networks often relies on large-scale labeled data, which is extremely expensive to acquire. Active Learning (AL) aims to reduce labeling costs by selecting the most informative samples for labeling. However, existing active learning research largely focuses on single-domain scenarios. In multi-domain scenarios, data distribution, task complexity, and learning dynamics differ significantly across domains.

[0003] The effectiveness of active learning strategies in existing methods depends on the budget-regime dependent phase. Diverse strategies are more effective in the early exploration phase, while uncertain strategies are better in the later exploitation phase. In multi-domain environments, simple domains may quickly enter the exploitation phase, while difficult domains are still in the exploration phase, making globally fixed strategies unsuitable for all domains. Furthermore, existing methods either only optimize budget allocation but use fixed strategies (such as CAL methods) or only optimize strategy selection but ignore cross-domain budget competition (such as traditional Automated AL). This decoupled design leads to budget shortsightedness, meaning that the budget allocated to a domain may not match its current optimal strategy, resulting in wasted resources. Summary of the Invention

[0004] To address the shortcomings mentioned in the background section, the present invention aims to provide a method and system for joint strategy selection and budget allocation across multiple domains. Firstly, the objective of this invention can be achieved through the following technical solution: a multi-domain joint strategy selection and budget allocation method, the method comprising the following steps: Obtain unlabeled sample pools pre-divided into multiple heterogeneous domains and the historical learning gain of each heterogeneous domain. Perform feedback-driven domain budget allocation calculation based on the historical learning gain of each heterogeneous domain to obtain the query sample quota for each domain. Based on the query sample quota for each domain, a strategy is selected using a preset Thompson sampling method to obtain the optimal strategy. Based on the optimal strategy, unlabeled sample pools from multiple heterogeneous domains are filtered to obtain candidate samples. The budget-aware composite reward for candidate samples is calculated based on a preset gating coefficient, wherein the composite reward is calculated based on uncertainty and representativeness weights.

[0005] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of calculating the feedback-driven domain budget allocation based on the historical learning gain of each heterogeneous domain, as follows: Computational domain In the Smoothing utility estimation of wheels : in, Let be the average compound reward achieved in round t. This is a momentum factor used to balance historical trends with current feedback; Calculate the original budget preference vector The final budget ratio is obtained by correcting for non-regression constraints. : in, Obtained by smoothing utility normalization, For domain The current cumulative annotation ratio; The maximum balance method is used to map continuous budget proportions to discrete integer sample quotas. .

[0006] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of selecting a strategy based on the query sample quota for each domain through a preset Thompson sampling, as follows: The strategy selection is based on Thompson sampling. The Beta posterior distribution of each candidate strategy is maintained independently for each domain. The optimal active learning strategy for the current round is adaptively selected through probability sampling to adapt to the asynchronous learning stages of different domains.

[0007] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the strategy selection process based on Thompson sampling includes: Modeling strategy selection as a multi-armed slot machine problem, for the domain Strategies in Maintain its posterior probability that follows a Beta distribution. ,in and respectively strategy Success and failure counts in this field; In each round, samples are taken from the posterior distribution. The strategy with the largest sample value is selected as the candidate strategy. A hysteresis switching mechanism is introduced to maintain training stability, and the switching only occurs when the expected return of the new candidate policy exceeds the expected return of the current policy by a threshold. Or, satisfy the preset random switching probability. Only when the time is right will a strategy switch be executed; otherwise, the strategy used in the previous round will be retained.

[0008] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the filtering of unlabeled sample pools of multiple heterogeneous domains based on the optimal strategy is based on diversity-aware filtering, and when the diversity-aware filtering is based on the currently selected optimal strategy as an uncertain sampling strategy, filtering is performed first and then clustering.

[0009] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the diversity-aware filtering process is as follows: First, an expansion candidate pool with a size equal to a multiple of the budget is retrieved based on the strategy score. The number of expansions... ,in For expansion factor; K-Means clustering is performed on the sample features in the expanded candidate pool to calculate cluster centers. The nearest neighbor samples of each cluster center are selected to form the final query batch, thereby maximizing feature coverage within the batch and eliminating information redundancy. In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of calculating the budget-aware composite reward of candidate samples based on a preset gating coefficient, as follows: Calculate samples Uncertainty score and representative score ,in Based on edge sampling calculation and batch-level normalization Cosine similarity calculation based on sample features and domain prototype; Constructing the gating coefficient Its proportion of the domain budget and training progress Joint decision: in, For the Sigmoid function, For budget sensitivity slope, This is the budget baseline offset. The time adjustment function is linearly increasing over time; when the domain budget is sufficient, the gating coefficient increases, and the reward calculation tends to utilize it; when the budget is scarce, the gating coefficient decreases, and the reward calculation tends to explore it. Calculate the final budget-perceived compound reward r(x): in, This indicates a normalization operation. This is a diversity scaling factor that decays as training progresses.

[0010] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the budget-aware composite reward is used to feed back and update the utility estimation parameters of the budget allocator and the posterior distribution parameters of the policy router for the next iteration, as follows: After obtaining manually labeled samples, the true cross-entropy loss of the samples is calculated, and the realized reward is generated based on the ranking of this loss within the batch. ,use Update the Beta posterior distribution parameters for the corresponding strategy: increase the success count if the reward is high. Otherwise, increase the failure count. At the same time Update the smoothing utility estimate for the corresponding domain. This allows for dynamic adjustment of the domain's budget allocation in the next round.

[0011] Secondly, in order to achieve the above objectives, this invention discloses a multi-domain joint strategy selection and budget allocation system, comprising: The data acquisition module is used to acquire unlabeled sample pools that are pre-divided into multiple heterogeneous domains and the historical learning gain of each heterogeneous domain. Based on the historical learning gain of each heterogeneous domain, feedback-driven domain budget allocation calculation is performed to obtain the query sample quota for each domain. The data processing module is used to select the optimal strategy based on the query sample quota of each domain through preset Thompson sampling, and then filter the unlabeled sample pools of multiple heterogeneous domains based on the optimal strategy to obtain candidate samples. The allocation output module is used to calculate the budget-aware composite reward of candidate samples based on a preset gating coefficient, wherein the composite reward is calculated based on uncertainty and representativeness weight.

[0012] The beneficial effects of this invention are: This invention incorporates multi-domain budget allocation and strategy selection into a unified optimization framework, employs a lightweight Thompson sampling strategy for routing, which can accurately capture the changes in the learning stages of different domains. While achieving rapid convergence in simple domains, it ensures sufficient exploration of difficult domains through diverse strategies. It designs a composite reward of non-regression budget constraints and budget awareness, achieving higher classification accuracy and lower annotation costs. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the principle architecture of the AUTOAL-ADAPT framework of this invention; Figure 3 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Example 1: like Figure 1 As shown, a multi-domain joint strategy selection and budget allocation method is proposed, which includes the following steps: S101: Obtain the unlabeled sample pools pre-divided into multiple heterogeneous domains and the historical learning gain of each heterogeneous domain. Based on the historical learning gain of each heterogeneous domain, perform feedback-driven domain budget allocation calculation to obtain the query sample quota for each domain. Specifically, by defining an unlabeled sample pool Divided into Heterogeneous domains Set the total number of rounds. and total budget for each round Initialize the shared model. and strategy pool The policy pool includes various active learning strategies such as entropy sampling, margin sampling, least confidence, K-Means clustering, BALD, BADGE, and free energy.

[0016] The process of calculating the feedback-driven domain budget allocation based on the historical learning gain of each heterogeneous domain is as follows: First, the smoothing utility estimate of the domain d in round t is calculated using the exponential moving average (EMA). The formula is as follows: in, The average compound reward obtained in the previous round for this domain (see step S104 for details). In a preferred embodiment, the momentum factor is... Set it to 0.7 to balance the impact of historical trends and current feedback.

[0017] Next, the original budget preference vector is calculated based on utility. : in, To prevent tiny constants with a denominator of zero.

[0018] To prevent hard domains from being allocated zero budget (i.e., "starved") due to poor initial performance, this invention introduces a non-regression constraint. Domain The current cumulative annotation ratio is ,in It is the cumulative number of labeled samples in this domain throughout history. This is the projected total budget up to the current round. The final budget allocation percentage. The calculation is as follows: This formula ensures that the current allocation ratio is not lower than its historical stock. Finally, the largest remainder method is used to adjust the continuous ratio. Convert to integer sample quota ,make sure .

[0019] S102: Based on the query sample quota of each domain, the strategy is selected through preset Thompson sampling to obtain the optimal strategy. Based on the optimal strategy, the unlabeled sample pools of multiple heterogeneous domains are screened to obtain candidate samples. The process of selecting a strategy based on the query sample quota for each domain through a preset Thompson sampling is as follows: The system performs policy selection based on Thompson sampling, independently maintains the Beta posterior distribution of each candidate policy for each domain, and adaptively selects the optimal active learning policy for the current round through probability sampling to adapt to the asynchronous learning stages of different domains.

[0020] The strategy selection process based on Thompson sampling includes: The strategy selection is modeled as a multi-armed slot machine problem, and the reward is treated as a Bernoulli process. For the domain Each strategy in Maintain its posterior probability, which follows a Beta distribution: in, and Representing strategies In the domain Success and failure counts are performed. To address non-stationary environments, a forgetting factor is introduced. (Preferably 0.9) Decrease the historical count.

[0021] In each round, samples are taken from the posterior distribution. The strategy with the largest sample value is selected as the candidate strategy. .

[0022] To avoid frequent policy oscillations, a hysteresis switching mechanism is introduced. The policy is switched from the previous round only when any of the following conditions are met. Switch to : in, This represents the expected value of the Beta distribution. The preferred threshold for switching is 0.02. To maintain the probability, it is preferable to set it to 0.6. Otherwise, maintain the strategy from the previous round. .

[0023] If the final selected strategy For uncertain sampling strategies (such as Entropy or Margin), diversity-aware execution is performed: first, samples of size [missing information] are retrieved. Expanded candidate pool (preferred expansion factor) Then, K-Means clustering is performed on the sample features in the expanded pool, and the nearest neighbor samples of each cluster center are selected to form the final query batch to eliminate information redundancy within the batch.

[0024] The selection of unlabeled sample pools from multiple heterogeneous domains is based on the optimal strategy. This is done using diversity-aware filtering. When the currently selected optimal strategy is an uncertain sampling strategy, the diversity-aware filtering performs filtering first and then clustering.

[0025] The diversity perception filtering process is as follows: First, an expansion candidate pool with a size equal to a multiple of the budget is retrieved based on the strategy score. The number of expansions... ,in For expansion factor; K-Means clustering is performed on the sample features in the expanded candidate pool to calculate the cluster centers. The nearest neighbor samples of each cluster center are selected to form the final query batch, so as to maximize the feature coverage within the batch and eliminate information redundancy.

[0026] S103: Calculate the budget-aware composite reward of candidate samples based on a preset gating coefficient, wherein the composite reward is calculated based on uncertainty and representativeness weight.

[0027] The process of calculating the budget-aware composite reward of candidate samples based on a preset gating coefficient is as follows: This step involves building a feedback signal for updating the system. The specific steps are as follows: First, calculate the two base scores for sample $x$: Uncertainty Score : Based on Margin Sampling calculation, and perform Rank-Normalization within the Batch to map it to the [0, 1] interval.

[0028] Representative score Calculate sample features With domain prototype The cosine similarity (to the mean of all labeled sample features in the current domain) is calculated using the following formula: .

[0029] To adapt to different budget states and training phases, a dynamic gating coefficient is constructed. : ) in, For the Sigmoid function; The slope for budget sensitivity is preferably set to 10; This is the budget baseline offset (usually set to a uniformly distributed value). This represents the current training progress. For the time adjustment function, linear scheduling is preferred. This causes the weights to gradually shift towards uncertainty as training progresses.

[0030] The final compound reward $r(x)$ is calculated as follows: in, This indicates a normalization operation; The diversity scaling factor decays over time, and the formula is as follows: The range decreased from 1.2 to 0.8.

[0031] Closed-loop update: After samples are labeled, the ranking based on the labeled real loss is used as the realized reward. The Beta distribution parameters are updated using the following formula to complete the loop closure: At the same time Feedback is provided to update the EMA utility estimate.

[0032] Experimental Verification and Performance Analysis: To verify the effectiveness of the proposed method, extensive comparative experiments were conducted on four standard multi-domain benchmark datasets: Office-Home, ImageCLEF, DomainNet-126, and PACS. As shown in Table 1, the proposed method (Ours) was compared with existing single-domain active learning methods (such as Entropy and BADGE) and multi-domain active learning methods (such as MultiAL and CAL). Experimental results show that, under the same annotation budget (e.g., N=200 or N=400), the proposed method achieved high classification accuracy on all datasets. Furthermore, the proposed method demonstrates a particularly significant advantage in the early stages of active learning with low budgets, proving that the feedback-driven budget control mechanism effectively prevents difficult domains from "starving" in the early stages.

[0033] Table 1 like Figure 2 As shown, the overall framework of AUTOAL-ADAPT in multi-domain active learning scenarios takes different data domains as input. First, it dynamically calculates the budget allocation ratio based on the historical learning gains of each domain, ensuring that annotation resources flow more towards domains that still have high learning gains or significant training needs. Then, within each domain, the system adaptively selects the most suitable sampling method from candidate active learning strategies through a ThompsonSampling-based policy selection module, and accordingly selects samples to be labeled. Simultaneously, the framework introduces a budget-aware composite reward mechanism, combining sample uncertainty, representativeness, and domain-level budget status to simultaneously guide policy selection and budget updates. The feedback results after labeling further update the domain-level gain estimate and policy posterior distribution, thus forming a closed-loop control process of "budget allocation—policy selection—sample query—reward feedback." Overall, Figure 2 The core idea of ​​the proposed method is emphasized: in multi-domain heterogeneous data environments, the allocation of labeling budget and the selection of active learning strategies are no longer treated separately, but rather the two are optimized collaboratively through a unified adaptive mechanism.

[0034] Example 2: To achieve the above objective, such as Figure 3 As shown, based on Embodiment 1, this invention discloses a multi-domain joint strategy selection and budget allocation system, including: The data acquisition module 11 is used to acquire unlabeled sample pools that are pre-divided into multiple heterogeneous domains and the historical learning gain of each heterogeneous domain. Based on the historical learning gain of each heterogeneous domain, feedback-driven domain budget allocation calculation is performed to obtain the query sample quota for each domain. Data processing module 12 is used to select the optimal strategy based on the query sample quota of each domain through preset Thompson sampling, and then filter the unlabeled sample pool of multiple heterogeneous domains based on the optimal strategy to obtain candidate samples. The allocation output module 13 is used to calculate the budget-aware composite reward of candidate samples based on a preset gating coefficient, wherein the composite reward is calculated based on uncertainty and representativeness weight.

[0035] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0036] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0037] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0038] The foregoing has shown and described the basic principles, main features, and advantages of this disclosure. Those skilled in the art should understand that this disclosure is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this disclosure. Various changes and modifications can be made to this disclosure without departing from its spirit and scope, and all such changes and modifications fall within the scope of this disclosure as claimed.

Claims

1. A multi-domain joint strategy selection and budget allocation method, characterized in that, The method includes the following steps: Obtain unlabeled sample pools pre-divided into multiple heterogeneous domains and the historical learning gain of each heterogeneous domain. Perform feedback-driven domain budget allocation calculation based on the historical learning gain of each heterogeneous domain to obtain the query sample quota for each domain. Based on the query sample quota for each domain, a strategy is selected using a preset Thompson sampling method to obtain the optimal strategy. Based on the optimal strategy, unlabeled sample pools from multiple heterogeneous domains are filtered to obtain candidate samples. The budget-aware composite reward for candidate samples is calculated based on a preset gating coefficient, wherein the composite reward is calculated based on uncertainty and representativeness weights.

2. The multi-domain joint strategy selection and budget allocation method according to claim 1, characterized in that, The process of calculating the feedback-driven domain budget allocation based on the historical learning gain of each heterogeneous domain is as follows: Computational domain In the Smoothing utility estimation of wheels : in, For the first The average compound reward achieved in each round, The momentum factor is used to balance historical trends with current feedback. Calculate the original budget preference vector The final budget ratio is obtained by correcting for non-regression constraints. : in, Obtained by normalization of smoothing utility. For domain The current cumulative annotation ratio; The maximum balance method is used to map continuous budget proportions to discrete integer sample quotas. .

3. The multi-domain joint strategy selection and budget allocation method according to claim 1, characterized in that, The process of selecting a strategy based on the query sample quota for each domain through a preset Thompson sampling is as follows: The strategy selection is based on Thompson sampling. The Beta posterior distribution of each candidate strategy is maintained independently for each domain. The optimal active learning strategy for the current round is adaptively selected through probability sampling to adapt to the asynchronous learning stages of different domains.

4. The multi-domain joint strategy selection and budget allocation method according to claim 3, characterized in that, The strategy selection process based on Thompson sampling includes: Modeling strategy selection as a multi-armed slot machine problem, for the domain Strategies in Maintain its posterior probability that follows a Beta distribution. ,in and respectively strategy Success and failure counts in this field; In each round, samples are taken from the posterior distribution. The strategy with the largest sample value is selected as the candidate strategy. A hysteresis switching mechanism is introduced to maintain training stability, and the switching only occurs when the expected return of the new candidate policy exceeds the expected return of the current policy by a threshold. Or meet the preset random switching probability. Only when the time is right will a strategy switch be executed; otherwise, the strategy used in the previous round will be retained.

5. The multi-domain joint strategy selection and budget allocation method according to claim 1, characterized in that, The process of filtering unlabeled sample pools from multiple heterogeneous domains based on the optimal strategy is based on diversity-aware filtering. When the currently selected optimal strategy is an uncertain sampling strategy, diversity-aware filtering is performed by first filtering and then clustering.

6. The multi-domain joint strategy selection and budget allocation method according to claim 5, characterized in that, The diversity perception filtering process is as follows: First, based on the strategy score, an expansion candidate pool with a size equal to a multiple of the budget is retrieved, and the number of expansions is determined. ,in For expansion factor; K-Means clustering is performed on the sample features in the expanded candidate pool to calculate the cluster centers. The nearest neighbor samples of each cluster center are selected to form the final query batch, so as to maximize the feature coverage within the batch and eliminate information redundancy.

7. The multi-domain joint strategy selection and budget allocation method according to claim 1, characterized in that, The process of calculating the budget-aware composite reward of candidate samples based on a preset gating coefficient is as follows: Calculate samples Uncertainty score and representative score ,in Based on edge sampling calculation and batch-level normalization Cosine similarity calculation based on sample features and domain prototype; Constructing the gating coefficient Its proportion of the domain budget and training progress Joint decision: in, For the Sigmoid function, For budget sensitivity slope, This is the budget baseline offset. The time adjustment function grows linearly with time; when the domain budget is sufficient, the gating coefficient increases, and the reward calculation tends to utilize it; when the budget is scarce, the gating coefficient decreases, and the reward calculation tends to explore it. Calculate the final budget-perceived compound reward r(x): in, This indicates a normalization operation. This is a diversity scaling factor that decays as training progresses.

8. The multi-domain joint strategy selection and budget allocation method according to claim 7, characterized in that, The budget-aware composite reward is used to feed back and update the utility estimation parameters of the budget allocator and the posterior distribution parameters of the policy router for the next iteration, as follows: After obtaining manually labeled samples, the true cross-entropy loss of the samples is calculated, and the realized reward is generated based on the ranking of this loss within the batch. ,use Update the Beta posterior distribution parameters for the corresponding strategy: increase the success count if the reward is high. Otherwise, increase the failure count. At the same time Update the smoothing utility estimate for the corresponding domain. This allows for dynamic adjustment of the domain's budget allocation in the next round.

9. A multi-domain joint strategy selection and budget allocation system, employing the multi-domain joint strategy selection and budget allocation method as described in any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to acquire unlabeled sample pools that are pre-divided into multiple heterogeneous domains and the historical learning gain of each heterogeneous domain. Based on the historical learning gain of each heterogeneous domain, feedback-driven domain budget allocation calculation is performed to obtain the query sample quota for each domain. The data processing module is used to select the optimal strategy based on the query sample quota of each domain through preset Thompson sampling, and then filter the unlabeled sample pools of multiple heterogeneous domains based on the optimal strategy to obtain candidate samples. The allocation output module is used to calculate the budget-aware composite reward of candidate samples based on a preset gating coefficient, wherein the composite reward is calculated based on uncertainty and representativeness weight.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on a processor. When the processor loads and executes the computer program, it employs the multi-domain joint strategy selection and budget allocation method as described in any one of claims 1 to 8.

Images