Outlier class inhibition based domain adversarial partial domain adaptation knowledge transfer method and system

Abstract

The application discloses a kind of based on field confrontation partial domain adaptation knowledge transfer method and system of outlying class inhibition, it is related to artificial intelligence and robot technical field.Knowledge transfer framework is constructed, including field discriminator and classifier;Feature is extracted from the data of source domain and target domain;The extracted feature is aligned by field discriminator through field confrontation, and the propagation of outlying class samples in source domain is inhibited using the outlying class inhibition function of classifier;After receiving observation input, the common features of source domain and target domain are calculated based on the feature output before action mapping layer, as the common knowledge of agent policy;According to the common features and the observation output of target agent, the transfer loss is calculated to update the policy network of target agent.The purpose is to solve the problem of poor adaptability of original strategy caused by change of task target point in mobile multi-agent same domain task scenario, and to improve the efficiency and accuracy of knowledge transfer through field confrontation alignment and outlying class inhibition dual mechanism.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and robotics, and more specifically to a domain-adaptive knowledge transfer method and system based on outlier suppression for domain adversarial purposes. Background Technology

[0002] With the widespread application of mobile multi-agent systems in complex task scenarios, such as logistics scheduling, disaster relief, and environmental monitoring, dynamic adjustment of agent task objectives has become a common requirement. However, this dynamic characteristic poses a severe challenge to agent policy learning and cooperative control: when task objectives change, directly reusing previously trained policy models not only fails to achieve the expected learning effect but may also lead to a significant decline in system performance. This manifests in two aspects: first, traditional multi-agent cooperative policy networks struggle to adapt to environmental changes caused by task adjustments, and the interaction relationships and policies between agents may become inefficient or even ineffective in new task scenarios; second, agents need to re-explore optimal policies rather than simply relying on the generalization ability of existing policies, which not only prolongs the training cycle but also significantly increases the computational cost during the task execution phase.

[0003] To address the aforementioned challenges, traditional multi-agent reinforcement learning algorithms struggle to achieve efficient responses to dynamic task objectives. Against this backdrop, introducing knowledge transfer methods to leverage the experience accumulated by agents in the source task to accelerate learning for new tasks has become an important research direction in this field. However, constrained by the inherent complexity of reinforcement learning algorithms, knowledge transfer still faces numerous technical bottlenecks. In computer vision, knowledge transfer typically relies on feature sharing mechanisms of image data; however, in reinforcement learning, the input of the policy network is deeply coupled with environmental characteristics, making it extremely sensitive to subtle changes in the task environment. This sensitivity makes it difficult to directly transfer pre-trained models to the target task in multi-agent reinforcement learning tasks. Although fine-tuning methods can achieve knowledge transfer to some extent by fixing some network layers and adjusting the remaining parameters, this method may compromise the optimality of the original policy, thereby reducing the transfer effect. Therefore, directly transferring policies often fails to achieve ideal results in multi-agent reinforcement learning tasks.

[0004] Another common transfer strategy is based on the knowledge extraction capabilities of neural networks, extracting high-value feature information from the source task for training the target task. This method accelerates the learning process and guides the model towards optimization that is more suitable for the target task by narrowing the exploration space of the target domain and avoiding blind searching. However, its core challenge lies in how to accurately extract the shared knowledge between the source and target tasks, ensuring the positive promotion of knowledge transfer and avoiding the risk of negative transfer. Although most domain adaptation methods based on domain adversarial learning networks promote knowledge transfer by learning domain-invariant features, their transfer efficiency is limited because they rely on the overlap of label spaces. In practical applications, the label space of the target domain is often a subset of the source domain. Therefore, Partial Domain Adaptation (PDA) methods are gradually showing unique advantages in reinforcement learning model transfer.

[0005] Partial domain adaptation (PDA) methods improve transfer performance by matching features of shared class samples between the source and target domains, effectively mitigating the negative impact of label space differences, and their effectiveness has been verified in multiple application scenarios. However, existing PDA methods still have a key drawback: outlier class samples may participate in the transfer process, leading to a risk of negative transfer. Therefore, how to solve the above-mentioned technical problems still requires further research by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a domain-adversarial partial domain-adaptive knowledge transfer method and system based on outlier suppression, which aims to solve the problem of poor adaptability of the original strategy due to changes in the task target point in mobile multi-agent task scenarios with the same domain. It improves the efficiency and accuracy of knowledge transfer through a dual mechanism of domain adversarial alignment and outlier suppression.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A domain adversarial partial domain adaptation knowledge transfer method based on outlier suppression includes the following steps:

[0009] Construct a knowledge transfer framework, including a domain discriminator and a classifier;

[0010] Select one transferable domain as the source domain and another transferable domain as the target domain, and extract features from the data in the source domain and the target domain;

[0011] The extracted features are adversarially aligned using a domain discriminator, and the outlier suppression function of the classifier is used to suppress the propagation of outlier samples in the source domain.

[0012] After receiving the observed input, the common features of the source and target domains are calculated based on the feature output before the action mapping layer, which serve as common knowledge for the agent's strategy.

[0013] The transfer loss is calculated based on common features and the observed output of the target agent, and used to update the policy network of the target agent.

[0014] Optionally, the extracted features are aligned adversarially using a domain discriminator. Specifically, the domain discriminator distinguishes samples from the source and target domains based on the extracted features through an adversarial learning mechanism. It also uses an entropy-based sample weighting strategy to assign different weights to shared class samples from the source and target domains and calculates the weights using an adversarial loss function to achieve cross-domain alignment of shared class samples.

[0015] Optionally, the outlier suppression function of the classifier can be used to suppress the propagation of outlier samples in the source domain. Specifically, the classifier performs label prediction on the target domain samples based on the extracted features, and assigns low prediction scores to outlier samples in the source domain through an adaptive weighted complementary entropy loss function to suppress the propagation from the source domain to the target domain, thereby ensuring the accuracy of knowledge transfer.

[0016] Optionally, a domain discriminator is used to distinguish between source and target domain samples, and the neural network aims to learn domain-invariant features, with the source and target domains having the same classification function.

[0017] Optionally, the detailed calculation formula for the domain discrimination loss of the domain discriminator to achieve adversarial alignment is as follows:

[0018]

[0019] in, It is the first in the source domain One sample, Source domain sample The tag, It is a weighting function based on entropy for relevant samples. It is a weight function based on sample labels. It's about the feature extractor network for the samples. yes Network parameters, It is a domain discriminator network. yes Network parameters, and These represent the number of samples in the source domain and the target domain, respectively. It is a trade-off parameter.

[0020] Optionally, a confidence-weighted complementary entropy objective loss function can be used, as shown in the following formula:

[0021]

[0022] in, It's about the feature extractor network for the samples. It is a classifier network. These are the network parameters of the feature extractor and the classifier, respectively. Indicates the number of samples in the source domain. It is the total number of categories. It is a label-based weighting function. It is the predicted probability vector output by the model. These are the true labels of the samples. The model predicts that the sample belongs to the true category. The probability, The model predicts that the sample belongs to another category. The probability, It's a hyperparameter. It is the number of categories of the actual labels in the target domain category. It is a normalized estimated class-level weight vector.

[0023] Optionally, the optimization objective of the domain adversarial learning network can be expressed as:

[0024]

[0025] in, These are the network parameters of the feature extractor, the neighborhood discriminator, and the classifier, respectively. It is the source domain classification cross-entropy loss function. It is the loss function that minimizes conditional entropy. It is a confidence-weighted complementary entropy objective loss function. It is a domain-discriminative loss function. , , It's about balancing hyperparameters.

[0026] A domain-adversarial partial-domain adaptive knowledge transfer system based on outlier suppression includes:

[0027] The framework building module is used to build knowledge transfer frameworks, including domain discriminators and classifiers;

[0028] The feature extraction module is used to select one transferable domain as the source domain and another transferable domain as the target domain, and extract features from the data in the source domain and the target domain.

[0029] The feature processing module is used to align the extracted features adversarially through the domain discriminator and to suppress the propagation of outlier samples in the source domain by utilizing the outlier suppression function of the classifier.

[0030] The network processing module is used to receive the observed input and calculate the common features of the source and target domains based on the feature output before the action mapping layer, as common knowledge of the agent's strategy.

[0031] The loss calculation module is used to calculate the transfer loss based on common features and the observation output of the target agent, and is used to update the policy network of the target agent.

[0032] As can be seen from the above technical solution, compared with the prior art, this invention discloses a domain-adversarial partial domain adaptation knowledge transfer method and system based on outlier suppression. It achieves adversarial alignment of shared class samples between the source and target domains by improving the domain discriminant loss function of the original Domain Adversarial Learning Network (DANN), and uses samples from the source domain to expand the target domain, thereby enhancing the efficiency of knowledge transfer. Then, a weighted complementary entropy objective function is designed to optimize the classification loss function in the original DANN, effectively suppressing the propagation of outlier samples and ensuring that the agent's policy learning is not interfered with by erroneous or irrelevant experiences during knowledge transfer. Finally, a task environment for verifying the performance of the transfer method is built using the Gym framework. A mobile multi-agent same-domain knowledge transfer task with only changes in the task objective point is designed, and the effectiveness of the proposed method is verified. Compared with the state-of-the-art methods, the OCS-DAPDA knowledge transfer method significantly improves the learning speed of the agent and enhances the performance of the final policy. The model exhibits the following core advantages in experiments:

[0033] 1. Dual mechanisms to improve transfer efficiency: Domain adversarial alignment aligns with the original samples in the source domain through a weighted sample strategy, alleviating domain offset and negative transfer problems; Outlier suppression suppresses source domain noise interference through an adaptive loss function, ensuring the accuracy of transferred knowledge.

[0034] 2. Low computational cost and strong generalization: It does not rely on generative models to synthesize target samples, reducing computational resource consumption; through progressive training strategies and the principle of entropy minimization, it improves the model's generalization ability under different task target point changing scenarios.

[0035] 3. Enhance feature discriminability: The weighted complementary entropy loss function enhances the inter-class margin, avoids source domain class confusion from propagating to the target domain, and solves the problem of insufficient feature discriminability in traditional domain adaptation methods.

[0036] This innovative approach effectively improves the efficiency of policy transfer and environmental adaptability of mobile multi-agent systems in the same domain task scenarios, providing a new solution for knowledge reuse in complex dynamic tasks. Attached Figure Description

[0037] 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, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of the network model of the present invention. Detailed Implementation

[0039] 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.

[0040] To address the negative transfer problem caused by asymmetric label distribution and outlier sample propagation in partial domain adaptation knowledge transfer, this invention proposes a domain adversarial partial domain adaptation knowledge transfer method based on outlier suppression. This method first uses a domain adversarial learning network to align shared class features between the source and target domains, extracting transfer knowledge adapted to the target task, and expanding the target domain data with source domain samples to improve transfer efficiency. Second, a weighted complementary entropy objective function is designed to optimize the classification loss function of a traditional deep adversarial network (DANN), suppressing the propagation of outlier samples and ensuring that the agent's learning is not interfered with by errors or irrelevant experience. Finally, this method is integrated into a mobile multi-agent collaborative same-domain knowledge transfer framework, forming an innovative knowledge transfer scheme.

[0041] Example 1

[0042] This embodiment discloses a domain adversarial partial domain adaptation knowledge transfer method based on outlier suppression, such as... Figure 1 As shown, it includes the following steps:

[0043] Construct a knowledge transfer framework, including a domain discriminator and a classifier;

[0044] Select one transferable domain as the source domain and another transferable domain as the target domain, and extract features from the data in the source domain and the target domain;

[0045] The extracted features are adversarially aligned using a domain discriminator, and the outlier suppression function of the classifier is used to suppress the propagation of outlier samples in the source domain.

[0046] After receiving the observed input, the common features of the source and target domains are calculated based on the feature output before the action mapping layer, which serve as common knowledge for the agent's strategy.

[0047] The transfer loss is calculated based on common features and the observed output of the target agent, and used to update the policy network of the target agent.

[0048] Furthermore, the extracted features are aligned adversarially using a domain discriminator. Specifically, the domain discriminator distinguishes samples from the source and target domains based on the extracted features through an adversarial learning mechanism. It also uses an entropy-based sample weighting strategy to assign different weights to shared class samples from the source and target domains and calculates the adversarial loss function to achieve cross-domain alignment of shared class samples.

[0049] The knowledge transfer framework constructed by the method in this embodiment aims to solve the problem of poor adaptability of the original strategy due to changes in the target point in mobile multi-agent cooperative navigation tasks. It follows the definition of unsupervised partial domain adaptation (PDA), with source domain samples as... The target domain sample is The target tag is a subset of the source tags, that is... Let C represent the total number of categories in the target sample.

[0050] Furthermore, the outlier suppression function of the classifier is used to suppress the propagation of outlier samples in the source domain. Specifically, the classifier performs label prediction on the target domain samples based on the extracted features, and assigns low prediction scores to outlier samples in the source domain through an adaptive weighted complementary entropy loss function, thereby suppressing the propagation from the source domain to the target domain and ensuring the accuracy of transferred knowledge.

[0051] Furthermore, the domain discriminator is used to distinguish between source and target domain samples. The neural network aims to learn domain-invariant features, and the classification functions for the source and target domains are the same. By introducing a domain adversarial discriminant, the distributional differences of shared class features between the source and target domains can be mitigated.

[0052] Furthermore, the detailed calculation formula for the domain discrimination loss of the domain discriminator to achieve adversarial alignment is as follows:

[0053]

[0054] in, It is the first in the source domain One sample, Source domain sample The tag, It is a weighting function based on entropy for relevant samples. It is a weight function based on sample labels. It's about the feature extractor network for the samples. yes Network parameters, It is a domain discriminator network. yes Network parameters, and These represent the number of samples in the source domain and the target domain, respectively. It is a trade-off parameter.

[0055] Furthermore, to enhance attention to outlier samples with lower confidence levels and cross-entropy loss, a confidence-weighted complementary entropy objective loss function is designed, as shown in the following formula:

[0056]

[0057] in, It's about the feature extractor network for the samples. It is a classifier network. These are the network parameters of the feature extractor and the classifier, respectively. Indicates the number of samples in the source domain. It is the total number of categories. It is a label-based weighting function. It is the predicted probability vector output by the model. These are the true labels of the samples. The model predicts that the sample belongs to the true category. The probability, The model predicts that the sample belongs to another category. The probability, It's a hyperparameter. It is the number of categories of the actual labels in the target domain category. It is a normalized estimated class-level weight vector. The confidence-weighted complementary entropy target loss utilizes adaptively weighted complementary entropy as a regularization term in the supplementary training strategy, which is more efficient.

[0058] Furthermore, integrating all the formulas, the optimization objective of the improved domain adversarial learning network can be expressed as:

[0059]

[0060] in, These are the parameters of the feature extractor, the neighborhood discriminator, and the classifier, respectively. It is the source domain classification cross-entropy loss function. It is the loss function that minimizes conditional entropy. It is a confidence-weighted complementary entropy objective loss function. It is a domain-discriminative loss function. , , It's about balancing hyperparameters.

[0061] This embodiment also discloses a domain adversarial partial domain adaptation knowledge transfer system based on outlier suppression, including:

[0062] The framework building module is used to build knowledge transfer frameworks, including domain discriminators and classifiers;

[0063] The feature extraction module is used to select one transferable domain as the source domain and another transferable domain as the target domain, and extract features from the data in the source domain and the target domain.

[0064] The feature processing module is used to align the extracted features adversarially through the domain discriminator and to suppress the propagation of outlier samples in the source domain by utilizing the outlier suppression function of the classifier.

[0065] The network processing module is used to receive the observed input and calculate the common features of the source and target domains based on the feature output before the action mapping layer, as common knowledge of the agent's strategy.

[0066] The loss calculation module is used to calculate the transfer loss based on common features and the observation output of the target agent, and is used to update the policy network of the target agent.

[0067] Example 2

[0068] A domain adversarial partial domain adaptive knowledge transfer (OCS-DAPDA) method based on outlier suppression is applied to a mobile multi-agent cooperative navigation task. The specific steps are as follows:

[0069] Step 1: The pre-trained multi-agent policy network under the source task (such as an agent navigating from point A to point B) is structurally adjusted and expanded. Then, the source domain sample data (including environmental observations, action records and corresponding labels for successfully reaching point B during the navigation process from point A to point B) and the target domain sample data (environmental observations during the agent's navigation from point A to point C, without labels) are fed into the feature extractor of the improved domain adversarial part of the domain adaptation network.

[0070] Step 2: The feature extractor extracts features from the source and target domain data to generate corresponding feature representations. Based on these features, the domain discriminator distinguishes between source and target domain samples through an adversarial learning mechanism. At the same time, using an entropy-based sample weighting strategy, different weights are assigned to shared class samples in the source and target domains. The improved adversarial loss function is used to calculate the weights, reducing the distribution differences between the source and target domains and achieving cross-domain alignment of shared class samples.

[0071] Step 3: The classifier performs label prediction on the target domain samples based on the extracted features. At the same time, through the adaptive weighted complementary entropy loss function, it assigns low prediction scores to outlier samples in the source domain to suppress their propagation to the target domain, prevent misclassification between different categories, and ensure the accuracy of transferred knowledge.

[0072] Step 4: After receiving the observation input from the target agent during its navigation from point A to point C, extract the feature output before the action mapping layer as common features; combine the common features with the observation output of the target agent to calculate the transfer loss, which is used to measure the adaptation difference between the source domain (navigation from A to B) knowledge and the target domain (navigation from A to C) task.

[0073] Step 5: Through the backpropagation mechanism, the policy network of the target agent is iteratively updated using the transfer loss. The domain-invariant features extracted from the source domain and the reliable knowledge after suppressing outliers are transferred to the target domain, thereby improving the adaptability of the agent in the task of navigating from point A to point C and completing the knowledge transfer.

[0074] Example 3

[0075] Based on knowledge transfer in mobile multi-agent material transportation tasks, this invention discloses a knowledge transfer framework based on OCS-DAPDA, applied to mobile multi-agent material transportation tasks, specifically including the following components and workflow:

[0076] Instruction Processing Module: Receives relevant data and instructions for the source task (the agent transports goods from warehouse X to station Y) and the target task (the agent transports goods from warehouse X to station Z). Feature Extraction Component: Inputs samples from the source domain (transportation data from warehouse X to station Y) and the target domain (transportation data from warehouse X to station Z) into the shared feature extractor to obtain feature representations.

[0077] Domain discriminant component: The domain discriminator is based on the output of the feature extractor and distinguishes between source domain and target domain samples through adversarial learning. It uses the original sample alignment strategy of the source domain to alleviate the distribution differences between domains.

[0078] Classification processing component: The classifier predicts the labels of samples in the target domain and uses an adaptive weighted complementary entropy loss function to suppress interference from outliers in the source domain, ensuring prediction accuracy.

[0079] Action Decision Module: Based on the common features output by the action mapping layer and combined with the observations of the target agent during the transportation process to station Z, the module calculates the migration loss, updates the target agent's policy network, and realizes the knowledge transfer of the transportation task from warehouse X to station Z.

[0080] Example 4

[0081] A knowledge transfer method based on OCS-DAPDA is applied to mobile multi-agent patrol tasks in complex environments, taking a forest patrol scenario as an example: The multimodal large model acquires instructions: the source task is for agents to patrol a region M along a fixed route; the target task is for agents to adjust their patrol route within region M based on unforeseen circumstances (such as fire warnings). The pre-trained multi-agent policy network under the source task (fixed route patrol) is structurally adjusted. In region M, source domain sample data (environmental images, terrain information, patrol path records, and corresponding normal patrol labels during fixed route patrols) and target domain sample data (including environmental images and terrain information for unforeseen circumstances such as fire warnings, without labels) are collected and fed into the feature extractor. The YOLO object detection algorithm is called to extract features from the environmental images of the source and target domains, generating feature representations. The domain discriminator distinguishes between source and target domain samples based on these features, using a conditional entropy term to encourage the model to generate high-confidence predictions for target domain samples. Simultaneously, the original source domain samples are used to align the inter-domain label distribution, reducing domain shift.

[0082] The classifier uses an adaptive weighted complementary entropy loss function to suppress outlier samples in the source domain that are unrelated to the sudden situation, preventing them from interfering with the target domain prediction; based on the extracted features, it performs label prediction on the target domain samples to determine the current patrol status and whether the route needs to be adjusted.

[0083] After receiving the observation input from the target agent during the patrol process, the feature output before the action mapping layer is extracted as the common feature; the migration loss is calculated by combining the common feature with the observation output of the target agent when dealing with emergencies such as fire warning.

[0084] By using the backpropagation mechanism, the policy network of the target agent is iteratively updated using the transfer loss. The domain-invariant features extracted from the source domain and the reliable knowledge after suppressing outliers are transferred to the target domain, enabling the agent to adjust the patrol route according to emergencies and complete the knowledge transfer of patrol tasks in complex environments.

[0085] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0086] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for domain-adversarial partial domain adaptation knowledge transfer based on out-of-class inhibition, characterized in that, The method comprises the following steps: Constructing a knowledge transfer framework including a domain discriminator and a classifier; Selecting a transferable domain as a source domain and another transferable domain as a target domain, and extracting features from data of the source domain and the target domain; Aligning the extracted features through the domain discriminator in an adversarial manner, and inhibiting the spread of out-of-class samples in the source domain by using the out-of-class inhibition function of the classifier; After receiving an observation input, calculating common features of the source domain and the target domain based on feature outputs before an action mapping layer, as common knowledge of an agent policy; Calculating a transfer loss based on the common features and an observation output of the target agent, and using the transfer loss to update a policy network of the target agent; Aligning the extracted features through the domain discriminator in an adversarial manner, specifically: the domain discriminator distinguishes samples of the source domain and the target domain based on the extracted features through an adversarial learning mechanism, and uses a sample weight strategy based on entropy to assign different weights to shared class samples of the source domain and the target domain, and calculates through an adversarial loss function to realize cross-domain alignment of the shared class samples; Inhibiting the spread of out-of-class samples in the source domain by using the out-of-class inhibition function of the classifier, specifically: the classifier performs label prediction on target domain samples based on the extracted features, and assigns low prediction scores to out-of-class samples in the source domain through an adaptive weighted complementary entropy loss function to inhibit the spread of the source domain to the target domain, so as to ensure the accuracy of the transferred knowledge.

2. The method of claim 1, wherein the method is based on out-of-distribution suppression. The domain discriminator is used to distinguish samples of the source domain and the target domain, and the neural network aims to learn domain-invariant features, and the classification functions of the source domain and the target domain are the same.

3. The method of claim 1, wherein the method is characterized by, The detailed calculation formula of the domain discrimination loss of the domain discriminator in the adversarial alignment is as follows: in, It is the first in the source domain One sample, Source domain sample The tag, It is a weighting function based on entropy for relevant samples. It is a weight function based on sample labels. It's about the feature extractor network for the samples. yes Network parameters, It is a domain discriminator network. yes Network parameters, and These represent the number of samples in the source domain and the target domain, respectively. It is a trade-off parameter.

4. The method of claim 1, wherein the method is based on out-of-distribution suppression. The target loss function based on confidence weighting is as follows: wherein, is a feature extractor network on the sample, is a classifier function, is a network parameter of is a network parameter of denotes the number of source domain samples, is the total number of classes, is a weight function based on the sample label, is the predicted probability vector of the model output, is the true label of the sample, is the probability that the model predicts the sample belongs to the true class , is the probability that the model predicts the sample belongs to other classes , is a hyperparameter, is the number of classes with true label in the target domain class, is the normalized estimated class-level weight vector.

5. The method of claim 1, wherein the method is based on out-of-distribution suppression. The optimization target of the domain adversarial learning network is as follows: wherein, are network parameters of the feature extractor, the domain discriminator, and the classifier, respectively, is a weighted source domain classification loss, is a target domain entropy minimization loss, is a weighted core loss, is a balanced domain adversarial loss, , , is a trade-off hyper-parameter.

6. A system for domain-adversarial partial domain adaptation knowledge transfer based on out-of-class inhibition, the system comprising: The method comprises the following steps: A framework construction module is configured to construct a knowledge transfer framework including a domain discriminator and a classifier; A feature extraction module is configured to select a transferable domain as a source domain and another transferable domain as a target domain, and extract features from data of the source domain and the target domain; A feature processing module is configured to align the extracted features through the domain discriminator in an adversarial manner, and inhibit the spread of out-of-class samples in the source domain by using the out-of-class inhibition function of the classifier; A network processing module is configured to, after receiving an observation input, calculate common features of the source domain and the target domain based on feature outputs before an action mapping layer, as common knowledge of an agent policy; A loss calculation module is configured to calculate a transfer loss based on the common features and an observation output of the target agent, and use the transfer loss to update a policy network of the target agent; The extracted features are aligned through the domain discriminator in an adversarial manner, specifically: the domain discriminator distinguishes samples of the source domain and the target domain based on the extracted features through an adversarial learning mechanism, and uses a sample weight strategy based on entropy to assign different weights to shared class samples of the source domain and the target domain, and calculates through an adversarial loss function to realize cross-domain alignment of the shared class samples; The outlier suppression function of the classifier is used to suppress the propagation of outlier samples in the source domain. Specifically, the classifier performs label prediction on the target domain samples based on the extracted features. Through an adaptive weighted complementary entropy loss function, outlier samples in the source domain are given low prediction scores to suppress the propagation from the source domain to the target domain, thereby ensuring the accuracy of knowledge transfer.

Images