A Multi-Agent Task Understanding Method Based on Extreme Learning Machine
By employing a multi-agent task understanding method based on extreme learning machines, and using situational information and capability values to train the network structure, the problem of inconsistent instruction understanding in multi-agent systems is solved, achieving accurate task understanding and collaborative combat.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING INST OF TECH
- Filing Date
- 2020-11-13
- Publication Date
- 2026-05-26
Smart Images
Figure CN112257875B_ABST
Abstract
Description
Technical Field
[0001] This disclosure belongs to the field of multi-agent task understanding technology, and in particular relates to a multi-agent task understanding method based on extreme learning machine. Background Technology
[0002] With the rapid development of computer technology, communication technology, and artificial intelligence technology, traditional warfare models and mindsets have also changed. As modernization and informatization advance, various new technologies are being applied to combat systems, and combat methods are constantly upgrading and evolving. Multi-agent systems, with their advantages of reduced casualties, high mobility and survivability, flexible configuration, and suitability for performing various dangerous missions in harsh environments, have become an important force on the new battlefield.
[0003] Deep integration of multi-agent systems with manned platforms is crucial for forming an integrated and collaborative system. The OODA loop model describes the four stages of decision-making: observation, judgment, decision, and action. Task understanding in multi-agent systems is a vital part of the OODA loop, referring to the unmanned platform's understanding of the environment and situation to produce a task understanding result that aligns with the insufficiently expressed or unclear instructions issued by the manned platform. Classic task understanding methods include hierarchical task decomposition and natural language-to-action instruction mapping. However, these methods rely to some extent on the subjectivity of expert systems, which can lead to different task understanding results for the same instructions under the same conditions. Summary of the Invention
[0004] In view of this, this disclosure proposes a multi-agent task understanding method based on extreme learning machine, which can make full use of situational information of the battlefield environment and information such as the ability values of multi-agents to generate understanding results that effectively match the commander's thinking. It avoids the subjective factors of relying on expert systems to a certain extent, so that the same instructions can obtain the same task understanding results under the same conditions, thus ensuring the accuracy of task understanding.
[0005] According to one aspect of this disclosure, a multi-agent task understanding method based on extreme learning machines is proposed, comprising:
[0006] Initialize the parameters of the multi-agent system and the environmental situational awareness information;
[0007] Based on the multi-agent task, the multi-agent parameters, and environmental situational awareness information, the extreme learning machine formulates multi-agent task understanding sample data.
[0008] The multi-agent task understanding network structure of the extreme learning machine is determined based on the multi-agent task understanding sample data.
[0009] The task understanding network structure of the Extreme Learning Machine is trained using the task sample data of the multi-agent system to obtain the task understanding model of the Extreme Learning Machine.
[0010] When the multi-agent receives a task instruction, it acquires the current environmental situation awareness information and the parameters of the multi-agent, and inputs the current environmental situation awareness information and the parameters of the multi-agent into the task understanding model of the extreme learning machine to obtain the task understanding result of the multi-agent.
[0011] In one possible implementation, the task understanding sample data includes task understanding data and task understanding label data, which are divided into task understanding training data and task understanding test data.
[0012] In one possible implementation, the multi-agent task understanding network structure of the Extreme Learning Machine is trained using the multi-agent task understanding sample data to obtain the multi-agent task understanding model of the Extreme Learning Machine, including:
[0013] The initial task understanding network structure of the multi-agent learning machine is obtained by training the multi-agent task understanding network structure of the extreme learning machine using the multi-agent task understanding training data.
[0014] The initial task understanding network structure of the multi-agent machine is tested using the multi-agent task understanding test data. If the performance of the multi-agent machine is satisfied, the initial task understanding network structure of the multi-agent machine is saved as the task understanding model of the multi-agent machine of the extreme learning machine; otherwise, the initial task understanding network structure of the multi-agent machine of the extreme learning machine is adjusted.
[0015] In one possible implementation, the multi-agent parameters include the number of multi-agents n and the capability value of each multi-agent, where n is a positive integer;
[0016] The environmental situation awareness information includes the threat level, quantity, and / or distribution density of target objects in the environment.
[0017] In one possible implementation, the task understanding sample data format of the multi-agent system is [x1,...,x...]. n The task understanding sample label data format is [o1,...,o...]. i ,...,o n ], where x i Let y1 be the capability value of the i-th agent in response to the task instruction, y2 be the threat level value of the target object perceived through environmental situational awareness, and o be the number or distribution density of the target object perceived through environmental situational awareness.i The value of is 0 or 1, 1≤i≤n, where i is a positive integer and i is the multi-agent number;
[0018] In one possible implementation, the multi-agent task understanding network structure includes an input layer, a hidden layer, and an output layer; wherein the number of nodes in the input layer is n+2, the number of nodes in the output layer is 1, and the number of nodes in the hidden layer is less than the number of sample data.
[0019] In one possible implementation, adjusting the initial task understanding network structure of the multi-agent system in the extreme learning machine is equivalent to adjusting the number of nodes in the hidden layers of the task understanding network structure.
[0020] In one possible implementation, the tasks of the multi-agent system include search and attack tasks.
[0021] In one possible implementation, the activator function of the hidden layer is the sigmoid function.
[0022] The process involves initializing the multi-agent parameters and environmental situational awareness information; formulating task understanding sample data for the multi-agents using the Extreme Learning Machine (ELM) based on their tasks, parameters, and situational awareness information; determining the task understanding network structure of the ELM based on this sample data; training the network structure using the sample data to obtain the ELM's task understanding model; and, upon receiving a task instruction, acquiring current environmental situational awareness information and multi-agent parameters, inputting these into the ELM's task understanding model to obtain the task understanding result. This approach fully utilizes battlefield situational information and multi-agent capability values to generate understanding results that effectively align with the commander's thinking, avoiding the subjective factors of relying on expert systems to some extent. This ensures that the same instructions yield the same task understanding result under the same conditions, guaranteeing the accuracy of task understanding.
[0023] Other features and aspects of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description
[0024] The accompanying drawings, which are included in and form part of this specification, illustrate exemplary embodiments, features, and aspects of this disclosure together with the specification and serve to explain the principles of this disclosure.
[0025] Figure 1A flowchart of a multi-agent task understanding method based on an extreme learning machine according to an embodiment of the present disclosure is shown;
[0026] Figure 2 A schematic diagram of a hypothetical scenario of a multi-agent task understanding method based on an extreme learning machine according to an embodiment of the present disclosure is shown.
[0027] Figure 3 A schematic diagram of hypothetical scenario region division is shown in a multi-agent task understanding method based on extreme learning machine according to an embodiment of the present disclosure.
[0028] Figure 4 A schematic diagram of a multi-agent search task understanding network based on an embodiment of the present disclosure is shown.
[0029] Figure 5 A schematic diagram of a multi-agent strike task understanding network based on an extreme learning machine according to an embodiment of the present disclosure is shown.
[0030] Figure 6 A schematic diagram of a multi-agent task understanding result based on an embodiment of the present disclosure is shown;
[0031] Figure 7 A schematic diagram of task allocation for multiple agents based on an embodiment of the present disclosure, according to an embodiment of the present disclosure, is shown. Detailed Implementation
[0032] Various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.
[0033] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.
[0034] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.
[0035] In complex battlefield environments, multi-agent systems can perform dangerous missions under harsh conditions, effectively reducing casualties. During combat, commanders derive control tactics based on situational and environmental information and issue corresponding operational commands to the multi-agent systems. For efficient collaborative operations between commanders and multi-agent systems, the ability of the multi-agent systems to reasonably understand the commander's commands is crucial. Task understanding input data has multi-source characteristics; to better utilize the information in the input data, an extreme learning machine is employed for task understanding reasoning.
[0036] Figure 1 A flowchart illustrating a multi-agent task understanding method based on an embodiment of the present disclosure, using an extreme learning machine, is shown. Figure 1 As shown, the method may include:
[0037] Step S1: Initialize the multi-agent parameters and environmental situation awareness information.
[0038] In a specific hypothetical scenario, the various capabilities and related parameters of multiple agents can be initialized to different values. For example, parameters such as the speed, detection capability, attack capability, and initial position of the multiple agents, as well as the initialization operation of the target object (enemy multiple agents), setting its capability value, position, and other parameters. After initialization is completed, the model representations corresponding to all agents will display their visual positions in the specific hypothetical scenario.
[0039] In one example, the multi-agent parameters can be parameters of our multi-agent team, which may include the number of multi-agent teams n and the capability values of the multi-agent teams, where n is a positive integer. The environmental situational awareness information can be parameters of the enemy multi-agent team, which may include the threat level, number, or distribution density of the enemy multi-agent teams (target objects) in the environment.
[0040] Among them, the capability value of an agent can be the range that the agent can search for or attack, centered on the agent.
[0041] Figure 2 A hypothetical scenario diagram of a multi-agent task understanding method based on an embodiment of the present disclosure is shown. Figure 3 A schematic diagram of the hypothetical scenario region division of a multi-agent task understanding method based on an embodiment of the present disclosure, according to an embodiment of the present disclosure, is shown.
[0042] like Figure 2As shown, a hypothetical environment scenario built in the AnyLogic simulation environment includes obstacles such as rivers, with a size of 10km * 6cm. The number of friendly agents is 15; for example, agent 1 has a capability of 0.395km, agent 2 has a capability of 0.522km, etc. The enemy multi-agent threat level (target object) is 3, and the number is 5, etc. Then, the hypothetical scenario is preprocessed, dividing the large scenario area into several task areas based on the relative positions of the multi-agents, such as... Figure 3 As shown, based on the relative positions of the multiple agents, it will be as follows: Figure 2 The hypothetical scenario shown is divided into eight small regions. The obstacles and enemy agent types within different regions may be the same or different. In this way, task understanding sample data for multi-agent systems can be created using the hypothetical scenario and the relevant parameters of our multi-agent systems and enemy multi-agent systems in the scenario.
[0043] Step S2: Formulate task understanding sample data for the multi-agent learning machine based on the task of the multi-agent, the parameters of the multi-agent, and the environmental situational awareness information.
[0044] The task understanding sample data was randomly divided into two parts: task understanding training data and task understanding test data, with the number of task understanding test data being less than the number of task understanding training data.
[0045] The task understanding sample data is divided into two parts: task understanding data and task understanding label data. Therefore, the task understanding sample data format includes the task understanding data format and the task understanding label data format.
[0046] In one example, the multi-agent task understanding sample data format is [x1,...,x n The multi-agent task understanding sample label data format is [o1,...,o...]. i ,...,o n ], where x i Let y1 be the capability value of the i-th agent in response to the task instruction, y2 be the threat level value of the target object perceived through environmental situational awareness, and o be the number or distribution density of the target object perceived through environmental situational awareness. i The value of is 0 or 1, 1≤i≤n, where i is a positive integer and i is the multi-agent number.
[0047] For different tasks, such as typical search and attack tasks, we can create sample data formats and task understanding label data formats that are suitable for different tasks and are required for training extreme learning machine models.
[0048] For example, for a search task, if the number of agents is 15, then the sample data format for the designed search task would be [x1...x 15 The label data format is a one-dimensional vector [o1...o2], y1, y2, ... 15 ].
[0049] Where x1, x2, etc. represent agent 1, agent 2, etc., for the search command specified in the search instructions, such as Figure 2 , Figure 3 The capability value of the task area is defined as follows: y1 is the threat level value within the search target area obtained through environmental situational awareness, and y2 is the distribution density of enemy agents within the search target area obtained through environmental situational awareness. The tag data vector is [o1...o...]. 15 The vector elements of [o1...o2] take values of 0 or 1, representing whether the corresponding agent performs the search task. For example, when the label data vector [o1...o2]... 15 When the value is [0,0,0,0,0,1,0,0,0,1,0,0,0,0,0], it means that the 6th and 10th agents will perform the search task, while other agents will not perform the search task.
[0050] Based on the search area defined in step 2, the stage of generating search task samples is carried out. The multi-agents are located in certain positions in the scene. When the commander issues the command for the search task area, the task understanding module of the multi-agents is triggered. The AnyLogic simulation system will automatically record the current capability value of each agent and the current environmental situational awareness information value according to the specified search task sample data format to generate a sample. The commander creates corresponding label data according to the search task label data format based on experience. The completed search task understanding sample data and label data are stored in a text file. For example, 700 sets of search task understanding sample data are generated. 500 sets of search task understanding sample data are randomly selected to form the search task understanding sample training set, and the remaining 200 sets of search task understanding sample data form the search task understanding sample test set.
[0051] For strike missions, the number of multi-agents remains 15. Therefore, the sample data format for the designed strike mission is [z1,...,z]. 15 The label data format is a one-dimensional vector [O1,...,O2,k1,k2], where k1,k2,k ... 15 ].
[0052] Among them, z1, z 15 Agents such as Agent 1, Agent 2, etc., attack each other. Figure 2 , Figure 3The capability values of the target objects (enemy agents) in the task area are defined as follows: k1 is the sum of the threat values of the target object's group as perceived by the environmental situation, and k2 is the number of groups to which the target object belongs, as perceived by the environmental situation. The tag data vector is [O1,...,O...]. 15 The vector elements of [O1,...,O] take values of 0 or 1, representing whether the corresponding agent is assigned to perform the strike mission. For example, when the sample label data vector of the strike mission is [O1,...,O], the vector elements of [O1,...,O] take values of 0 or 1, representing whether the corresponding agent is assigned to perform the strike mission. 15 When the value is [0,0,0,0,0,1,0,0,0,1,0,0,0,0,0], it means that the 6th and 10th agents are assigned to carry out the strike mission, and the other agents do not carry out the strike mission.
[0053] After the search task and sample data generation are completed, the implementation example is run again to observe the results. Figure 2 , Figure 3 In the given environment, agents from a swarm of enemy agents are randomly selected as targets for the strike mission. When the commander issues the strike mission order, the AnyLogic simulation system automatically generates strike mission sample data according to the strike mission understanding sample data format. The commander then creates label data for the strike mission under this situation based on experience and according to the strike mission understanding sample label data format. After the labels are created, the strike mission understanding sample data and strike mission understanding sample label data are saved to the corresponding text files. Similar to the search mission, 700 sets of strike mission understanding sample data are generated. 500 sets are randomly selected to form the strike mission understanding sample training set, and the remaining 200 sets form the strike mission understanding sample test set.
[0054] Step S3: Determine the multi-agent task understanding network structure of the Extreme Learning Machine based on the multi-agent task understanding sample data.
[0055] In one example, the multi-agent task understanding network structure may include an input layer, a hidden layer, and an output layer; the number of nodes in the input layer is n+2 (n is the number of our multi-agents), the number of nodes in the output layer is 1, and the number of nodes in the hidden layer is less than the number of sample data.
[0056] The number of nodes in the input layer corresponds to the dimension of our multi-agent sample data, and the number of hidden layers can be set to a value less than the number of samples based on the commander's experience. The weight coefficients from the input layer to the hidden layer, and the bias coefficients of the hidden and output layers are initialized using random numbers, and the activation function of the hidden layers can be the sigmoid function.
[0057] Figure 4 , Figure 5Schematic diagrams of understanding networks for search and attack tasks of multi-agent systems based on extreme learning machines according to an embodiment of the present disclosure are shown respectively.
[0058] For example, utilizing our 15 intelligent agents and such Figure 2 , Figure 3 The hypothetical scenarios shown are used to train extreme learning machine models for search and attack tasks, respectively.
[0059] like Figure 4 As shown, for the Extreme Learning Machine (ELM) model of the training search task for the i-th agent, the number of input nodes in the multi-agent search task understanding network model structure is set to 17, the number of hidden layer nodes is 30, and the number of output layer nodes is 1. The biases a of the hidden layer neurons of this ELM are randomly initialized. j i) The bias b of the output neuron (i) The input weights W from the input layer to the hidden layer (i) The input weights β from the hidden layer to the output layer (i) The activation function is chosen as the sigmoid function. The principle of the Extreme Learning Machine (ELM) model for training the i-th agent on the attack task is the same as that for training the i-th agent on the search task, as follows: Figure 5 As shown, the input node of the Extreme Learning Machine Model for the strike mission will have an additional enemy unit group situation node.
[0060] Step S4: Use the task sample data of the multi-agent machine to train the multi-agent task understanding network structure of the extreme learning machine to obtain the multi-agent task understanding model of the extreme learning machine.
[0061] In one example, the task understanding network structure of the multi-agent learning machine can be trained using the multi-agent task understanding training data to obtain the initial task understanding network structure of the multi-agent learning machine; the initial task understanding network structure of the multi-agent learning machine can be tested using the multi-agent task understanding test data; if the performance of the multi-agent learning machine is satisfied, the initial task understanding network structure of the multi-agent learning machine is saved as the task understanding model of the multi-agent learning machine; otherwise, the initial task understanding network structure of the multi-agent learning machine is adjusted.
[0062] For our i-th agent, we train the multi-agent task understanding model of the extreme learning machine using the training sample set of search and attack tasks stored in the text, and test the performance of the multi-agent task understanding model using the predetermined test sample set of search and attack tasks. Taking the training of a multi-agent search task understanding model for an Extreme Learning Machine (ELM) as an example, the process involves loading 500 sets of search task understanding training samples corresponding to the i-th agent from the text, training the ELM's multi-agent search task understanding model, and then using the remaining 200 sets of search task understanding test samples to test the performance of the trained ELM's multi-agent search task understanding model. If the performance requirements for multi-agent tasks are met, the parameters of the ELM's multi-agent search task understanding model are saved; otherwise, the number of nodes in the hidden layer of the ELM's multi-agent search task understanding model is adjusted, and the above steps are repeated to generate a certain amount of new sample data for the ELM's multi-agent search task understanding model. This process of training and testing the ELM's multi-agent search task understanding model continues until a high-performance ELM multi-agent search task understanding model is obtained, and the parameters of the ELM's multi-agent search task understanding model are saved. The training process for the ELM's multi-agent attack task understanding model is similar to that of the search task. If an ideal Extreme Learning Machine (ELM) model cannot be obtained using the current samples, a certain number of search task samples can be regenerated and replaced with an equal number of samples in the current training set to retrain the multi-agent strike task understanding model of the ELM.
[0063] Step S5: After the multi-agent receives the task instruction, it acquires the current environmental situation awareness information and the multi-agent parameters, and inputs the current environmental situation awareness information and the multi-agent parameters into the multi-agent task understanding model of the extreme learning machine to obtain the task understanding result of the multi-agent.
[0064] Figure 6 A schematic diagram of a multi-agent task understanding result based on an embodiment of the present disclosure is shown; Figure 7 A schematic diagram of task allocation for multiple agents based on an embodiment of the present disclosure, according to an embodiment of the present disclosure, is shown.
[0065] After the multi-agent search task understanding model and strike task understanding model of the Extreme Learning Machine (ELM) are trained, each agent loads the corresponding trained ELM model. When the commander issues a search or strike task instruction to the agent, the AnyLogic simulation system obtains multi-source data such as current environmental situational awareness information and the capability values of friendly agents from the environment and inputs it into the multi-agent search task understanding model or strike understanding model framework of the ELM. Through inference and calculation, the task understanding result of the multi-agent is obtained as a one-dimensional vector with dimension n, where n is the number of friendly agents. Each element of the vector takes the value 0 or 1, corresponding to the task understanding result of each agent. Figure 6 As shown, when the multi-agent task understanding module of the Extreme Learning Machine receives the trigger command for the search task, the 15 multi-agents in the current environment are friendly unit 0, friendly unit 1...friendly unit 14. The search capability values of the 15 multi-agents are the values in the capability / distance column; the current environmental situational awareness information (threat level and distribution density of enemy multi-agents) are the values corresponding to the enemy threat level and enemy distribution density. The multi-agent search task understanding model of the Extreme Learning Machine obtains the result of the search task understanding through inference. Figure 7 As shown, the array [0,0,0,0,0,1,0,0,0,1,0,0,0,0,0] represents the search task assigned to the 6th and 10th agents, whose understanding results reasonably conform to the commander's intentions.
[0066] The process involves initializing the multi-agent parameters and environmental situational awareness information; formulating task understanding sample data for the multi-agents using the Extreme Learning Machine (ELM) based on their tasks, parameters, and situational awareness information; determining the task understanding network structure of the ELM based on this sample data; training the network structure using the sample data to obtain the ELM's task understanding model; and, upon receiving a task instruction, acquiring current environmental situational awareness information and multi-agent parameters, inputting these into the ELM's task understanding model to obtain the task understanding result. This approach fully utilizes battlefield situational information and multi-agent capability values to generate understanding results that effectively align with the commander's thinking, avoiding the subjective factors of relying on expert systems to some extent. This ensures that the same instructions yield the same task understanding result under the same conditions, guaranteeing the accuracy of task understanding.
[0067] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.
Claims
1. A multi-agent task understanding method based on Extreme Learning Machine, characterized in that, The method includes: Step S1: Initialize multi-agent parameters and environmental situational awareness information; The multi-agent parameters include the number of multi-agents n and the capability values of the multi-agents, where n is a positive integer; The environmental situational awareness information consists of parameters of the enemy multi-agent, including the threat level, number, or distribution density of the enemy multi-agent in the environment. The capability value of the intelligent agent is the range that the intelligent agent can search for or attack, centered on the intelligent agent. Step S2: Formulate task understanding sample data for the multi-agent learning machine based on the task of the multi-agent, the parameters of the multi-agent, and environmental situational awareness information; Among them, the task understanding sample data is randomly divided into two parts: task understanding training data and task understanding test data, with the number of task understanding test data being less than the number of task understanding training data. The task understanding sample data is divided into two parts: task understanding data and task understanding label data. Therefore, the task understanding sample data format includes the task understanding data format and the task understanding label data format. The data format for multi-agent task understanding is as follows: ; Multi-agent task understanding label data format is ; in, Let be the capability value of the i-th agent in response to the task instructions. This represents the threat level of the target object as perceived through environmental situational awareness. This refers to the number or distribution density of target objects detected through environmental situational awareness. The value is 0 or 1, 1≤i≤n, where i is a positive integer, i is the multi-agent number, and n is the number of our multi-agents; This is used to represent whether the i-th agent performs the task; Step S3: Determine the multi-agent task understanding network structure of the extreme learning machine based on the multi-agent task understanding sample data; The multi-agent task understanding network structure includes an input layer, a hidden layer, and an output layer; the number of nodes in the input layer is n+2, the number of nodes in the output layer is 1, and the number of nodes in the hidden layer is less than the number of sample data. The number of nodes in the input layer corresponds to the dimension of the sample data of our multi-agent system. The number of hidden layers is a value less than the number of samples. The weight coefficients from the input layer to the hidden layer are initialized with random numbers. The bias coefficients of the hidden layer and the output layer are initialized with random numbers. The activation function of the hidden layer is the sigmoid function. Step S4: Use the multi-agent task understanding sample data to train the multi-agent task understanding network structure of the Extreme Learning Machine to obtain the multi-agent task understanding model of the Extreme Learning Machine. Step S5: After the multi-agent receives the task instruction, it acquires the current environmental situation awareness information and the parameters of the multi-agent, and inputs the current environmental situation awareness information and the parameters of the multi-agent into the task understanding model of the extreme learning machine to obtain the task understanding result label data vector of the multi-agent.