A distributed computing method for collaborative situational awareness of intelligent agent clusters

Abstract

The present invention discloses a distributed computing method for collaborative situation perception of a cluster of intelligent agents. The algorithm includes information source confirmation and information dissemination and coordination methods; dynamic construction of attribute state values ​​that stimulate dissemination based on events and their types; and design of a hybrid modulation process for mutual influence between attribute states. The essence of collaborative situation perception of a group is a process of dimensionality reduction, transmission, and restoration of surrounding environmental information. Individuals must first determine the source of battlefield dynamic information, screen the main information, describe emergencies using dimensionality reduction and compression methods, and then transmit and cross-validate information through the local network structure of the cluster. Finally, the received dimensionality reduction information is used to re-evaluate and draw their own information map. The algorithm of the present invention is similar to the mechanism of biological hormone regulation and neural information transmission, and provides a computing method that can reduce the dimensionality of complex situation information and local coordination of intelligent agents under regional denial and low communication conditions.

Description

Technical Field

[0001] The present invention relates to the field of intelligent agent group collaboration, and specifically demonstrates a distributed computing and interaction method based on large-scale intelligent agent cluster situation collaborative perception. Background Art

[0002] Unmanned swarm combat is a "machine-to-machine collaboration" combat method. Human commanders can give instructions to the swarm as a whole without directly controlling each individual moving unit in the swarm. Make full use of the concept of distributed combat and adopt low-cost, reusable, miniaturized aerial combat platforms that can carry modular mission payloads such as reconnaissance, which are carried by carrier aircraft to the vicinity of the war zone for launch and recovery, greatly reducing the cost of combat units and improving combat flexibility. In particular, the self-organizing and flexible switching of unmanned swarm configurations will be a disruptive group intelligence emergence technology, containing a higher level of intelligence with rich new functions, strong functions and new attributes.

[0003] Unmanned swarm combat relies heavily on the sharing of environmental information between individuals, especially the collaborative awareness technology of swarm situation. In a complex and restricted battlefield environment, how to complete complex group situation awareness and collaborative tasks with low communication levels is an important research topic in drone swarm combat in various countries. Summary of the invention

[0004] According to the background introduction, the collaborative combat scenario in a denial environment requires the ability of local networking distributed computing for the intelligent agent cluster. At the same time, it must be able to restore and share the battlefield situation information as much as possible under low channel communication, and on this basis, form a dynamic collaborative combat strategy based on the current situation information. Due to the lack of information sharing and command transmission from the backend command center, the front-end intelligent agent cluster can only perform combat tasks autonomously through self-organization combined with the initial set plan, and correspondingly, it must spontaneously achieve coordination and cooperation between groups. This kind of cooperation requires the intelligent agent to complete the transmission of low-dimensional data and restore the overall situation information under low communication conditions in a combat environment where comprehensive information is incomplete and communication capabilities are challenged. In the management of the intelligent agent cluster, it is also necessary to consider whether it can reduce the commander's mental consumption, give full play to the ability of the cluster's intelligent autonomous combat, and achieve different combat style changes of the intelligent agent cluster through fewer parameter settings. To achieve the above capabilities, the scheme adopted by this algorithm is as follows:

[0005] The entire intelligent agent cluster situation collaborative perception algorithm includes three main contents: confirmation, sharing and selection of information sources; state attributes that stimulate propagation based on events and their types; and a mixed modulation process in which states influence each other.

[0006] A. Information source situation sharing

[0007] Step 1: Establish the agent’s own local database, including individual information maps for labeling and coordination, in order to transfer data.

[0008] Using R i Represents the basic information of individual i, including the location of the individual speed And other state representation quantities corresponding to event types (Right now For R i ), k represents the event and state type. The agent cluster has marked all individuals in the system during the initial deployment, so the agents can recognize each other and determine their unique numbers.

[0009] Step 2: Mark the information source of the newly discovered target.

[0010] When an agent discovers a high-value target or an interception unit of the other party during operation, the information source of the event, that is, the discovered target, will be marked as a node of state information in the agent's information map. For agent i, its basic state is marked as R i，m , array R i，m Contains all information after all i individuals detect and virtually evaluate target m (in addition to information such as position and speed, it also includes whether it is damaged, the amount of remaining ammunition, etc.). It represents the virtual information source that is marked by other individuals and transmitted to i, and m is its number. i，m and They share the same numbering system, and the “*” simply represents fuzzy information that is input by other individuals.

[0011] Step 3: Intelligent agents use communication functions to conduct local networking and data transmission, and number the information in different forms according to the source of the information.

[0012] When individual i and individual j meet and they can realize local networking for information transmission, agent i will transfer the information R of information source m to i，m Transmit and share to j. If j has not observed m before, the information source will be marked as n is the number that marks the fuzzy information source in the information map of agent j.

[0013] Step 4: By introducing the first marking time of the information source, coordinate and unify the numbering order of the information sources within the agent cluster.

[0014] In order to ensure the coordination requirements between agents, the target mark number m of observer i should be transmitted to receiver j, so that the local cluster will process the same number of objects when gathering and attacking.i，m and They share the same numbering system, so whether they are discovered by themselves or transmitted by other agents, the number of the target will continue as the number of tags increases. Assume that individuals i and j are at t 1 and t 2 The information sources a and b are observed at time t and the message is passed to individual k, but there is no network between i and j. 1 At time i, the information R about information source a i，m Passed to k, marked as In t 2 At this moment, j and k come into contact and form a network. j collects the R j，m Passed to k, it is possible that number m has already been used. The time stamp in the information is t 1 , R j，m The time stamp in the information is t 2 , when j and k exchange information, the information transmitted by j should be placed at the end, so it is marked as The message sent by k should be placed in the R of j's information map list. j，m In front of, so R j，m+1 =R j，m (All numbers after m are automatically increased by one),

[0015] Step 5: Solve the problem of duplicate labeling of a single source by merging similar information.

[0016] When individuals i and j discover the same unit, the process of transmitting information to k requires checking the input information. Represents the location information of the information source. When the two information sources are considered to be the same, the sum can be taken to get the mean, where r vague is the blur radius. Of course, in addition to integrating information about static targets based on position information, we can also integrate information about moving targets based on speed information. In addition, the newly input virtual information source It also needs all the information of its own mark R j，m For comparison, if The input information is ignored, and only the information source observed by itself is retained and recorded. Through the above operation, the event source in the active event can be marked on the map information of the intelligent body in the form of a virtual information source and transmitted and shared, as if the information of these event sources is uniformly distributed by the background.

[0017] B event situation sharing

[0018] The local situation information of the battlefield can be represented as a collection and continuation of a series of emergencies (such as being attacked, discovering targets, cruising to new areas, attacking targets, etc.). The intelligent agents involved in the events can generate attribute state representation values ​​corresponding to the type of events according to the type and nature of the events in the event activities, and at the same time save and record the most basic description of the events. Then, through the short-distance local networking between the intelligent agents, small data transmission and exchange can be achieved to share the local situation changes caused by the current events with the entire cluster. On this basis, different attribute state representation values ​​are used to mix and modulate the response actions that the intelligent agents should currently take, so as to realize the function of group collaborative combat.

[0019] Step 1: The intelligent individual records the local battlefield situation information through specific events and the corresponding characteristic state values ​​based on them.

[0020] Corresponding to the occurrence of an event, agent i will self-stimulate the corresponding state value Where k is the event type.

[0021] Step 2: Create any feature state for all event types Differential equation evolving over time:

[0022]

[0023] The three items on the right correspond to the self-excitation of the state value caused by the occurrence of the event, the conduction of the signal source state, and the decay of the state over time, where α, β, and γ are the mixed modulation coefficients, and δ(t event -t) represents the occurrence process of the event at time t.

[0024] Step 3: According to different event types, establish a status information transmission mode based on characteristic status.

[0025] The conduction process of the signal source state is

[0026]

[0027] C i represents the set of all information sources that can interact with individual i, and G is the conduction modulation function. The individual state value S that is affected and changed according to event type k k In the propagation interaction, it will be affected by other information sources about the state type, that is, the state propagation process will be affected by the numerous information sources R j Extract And other information to interact.

[0028] Step 4: Based on the event type, distinguish the events into active events and passive events.

[0029] Different event types have different roles played by various parts in the differential equation of a state change of an individual. According to the task characteristics of a large-scale intelligent agent cluster, the state attributes caused by its event type can be roughly divided into passive events such as the "fear" state stimulated by the injury event, the "excitement" state obtained by unimpeded task execution, and the "desire" state stimulated by the detection and approach behavior of a specific target, and the "disgust" state stimulated by a specific interception or expulsion event and the "disgust" state stimulated by the interfering individual.

[0030] Step 5: For source events, the state attributes need to further mark the state source.

[0031] Active events need to generate corresponding state representations according to the event source, that is, for the discovered high-value targets and their virtual information sources R i，m , the desire state of individual i should be It is used to distinguish its desire state for other high-value targets. Through the process of stimulating the desire state of the cluster by different information source targets and the propagation between agents, it is possible to determine the information source with the greatest attraction for a single agent as the current task target. h Similar to the Desire state, its activation process is based on the state of the enemy individuals that have been discovered to track and attack itself, so it is also a response to active events involving specific units.

[0032] C Mixed Modulation Control

[0033] Different attribute state values ​​follow the design of three steps: excitation, propagation, and decay, and are affected by the response event and the surrounding neighbor state. The attribute state value and the parameter adjustment of the related dynamic equation will be affected by other attributes. For example, a highly excited individual is likely to ignore the current threat situation and take some radical actions, so the excited state S e will suppress the fear state, making the decay modulation coefficient of the fear state γ d With S e is enhanced by the enhancement.

[0034] Step 1: When individual i finds the value target l, his desire state It will stimulate.

[0035] This property will adjust its coefficient α by the distance d desire , α desire =e -d By comparing the state attributes of each target You can sort multiple targets by their desired states and finally select target m.

[0036] Step 2: The individual is attracted by the target unit. The desire state drives the individual to approach the unit with the highest desire priority. The attraction is expressed as:

[0037]

[0038] c a is the intensity, r a is the attack range of the agent, is the unit direction vector from the agent to the goal.

[0039] Step 3: When the agent is attacked, its fear state S f It will stimulate.

[0040] The excitement state will suppress the fear value to a certain extent, and its value is adjusted to The mixed modulation process of excitation, propagation and decay is:

[0041]

[0042] The amount of fear value aroused will be proportional to the amount of damage the unit receives. * / D is related to this state. If the individual is continuously attacked over a period of time, this state will continue to be activated.

[0043] Step 4: Fear-driven individuals gain virtual volume expansion in the judgment of other individuals.

[0044] The damage event means that the fear unit and its surrounding space are in danger, so other agents will enlarge the volume of the units that are actually attacked based on the collision avoidance mechanism to avoid being attacked. The dynamic expression of avoidance is

[0045]

[0046] c d is intensity, projection is the unit vector in the vertical direction from the individual to the target The unit vector of the relative position of individual j to individual i The projection between.

[0047] Step 5: When the agent is not affected and can successfully perform the task, its excited state will continue to be excited.

[0048] If we do not consider the transmission between individuals, then When the state is stable, S e =α e / γ e Considering that the closer an individual is to the goal, the more excited he or she will be, the more likely he or she is to be excited. Let γe =1, then When an individual is in a "panic" state with a high fear value, the individual cannot obtain normal excitement from the surrounding neighbors, that is, fear will inhibit the interaction of excitement attributes between intelligent agents.

[0049] Step 6: When the agent is tracked, attacked, or disturbed by the opponent's discovered unit, the agent will feel disgusted with the disturbing unit.

[0050] The process of triggering the disgust state is related to the threat posed by the interference unit to the agent. The most intuitive relationship is the spatial distance. is the aversion escape distance constant of the interferer source l. The closer the interference unit is to the agent, the greater the threat. Correspondingly, the self-excitation effect can be established as where d * is the distance between the agent and the interferer.

[0051] Step 7: The expression of the aversion state for the agent dynamics is:

[0052]

[0053] G i is the set of interference units marked in the information map of i that can have a negative impact on itself, c h is the action intensity, projection is the unit vector in the direction of the relative position of the interference unit to the agent.

[0054] Step 8: The agent controls its velocity using the common self-driving dynamics equations.

[0055] Step 9: When the distance between the agents is less than the warning distance, the radial repulsion effect will be triggered.

[0056] Through the above three parts and steps, the transmission of information sources, the excitation of characteristic states of situation events, the mixed modulation control between different characteristic states and between dynamic models are completed when individuals respond to changes in environmental situations, thus achieving the goal of intelligent agent clusters from situation awareness collaboration to distributed computing and dynamic drive.

[0057] In general, to achieve situational awareness coordination and distributed information processing of intelligent clusters, it is necessary to first establish a labeling pattern of information and information sources that can be coordinated with each other, complete the unified identification of targets and the merging of similar labels. Then, by establishing a set of coordinated control and transmission mechanisms similar to biological hormone modulation and neural signal transmission in the local cluster, the complex battlefield information can be effectively simplified into the corresponding state information stored by each intelligent agent, and then transmitted and re-stimulated to other individuals. The intelligent agent can not only know its current state from the data transmitted by other individuals, but also adjust its own state based on it, and continue to transmit and activate these signals externally, which makes the information not only transmitted and shared between individuals, but also affects the action strategies of all intelligent agents. It is only necessary to make the dynamic equation of its action contain other attribute state values ​​in addition to the basic information to realize the control of individual behavior by state stimulation, transmission, and decay. On this basis, through the mutual activation and inhibition relationship between each attribute state, more and more complex interaction modes can be realized. Such a set of situational coordination algorithm systems that reduce the dimension of environmental information and regulate through mutual excitation and inhibition can help large-scale intelligent clusters to cope with complex war environments in a distributed intelligent self-organizing manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a schematic diagram of the content of the intelligent agent cluster situation collaborative perception algorithm;

[0059] Figure 2 A schematic diagram of the modulation relationship established according to four basic attribute states;

[0060] Figure 3 The collaborative process of numbering event sources in information source sharing;

[0061] Figure 4 The process of filtering and merging information sources in updating individual information maps;

[0062] Figure 5 Take a screenshot of the simulation process;

[0063] Figure 6 It is the result of the evolution of the average value of the survival status of the agent, the damage caused, and the proportion of agents in the attack circle over time;

[0064] Figure 7 The density distribution of the agent at different angles during the approach process with the value target as the center;

[0065] Figure 8 The distribution of the projection of the individual speed of the agent at different angles in the process of approaching the target in the direction from the individual to the target. DETAILED DESCRIPTION

[0066] The situational awareness, coordination and control process of intelligent agent clusters, such as Figure 1 As shown in the figure, it includes three parts: information source situation sharing, event situation sharing, and hybrid modulation control. The information source situation mainly refers to the process of describing the information source of the active event on the intelligent agent information map, sharing the creation process, and the collaborative process; the hybrid modulation control will mutually regulate according to the different states stimulated by the intelligent agent, which increases the complexity of its control level and can piece together a complex information landscape that is easy to control in a cluster from the various low-dimensional information transmitted by the surrounding neighbors; through the description and corresponding stimulation of various actual possible events, the intelligent agent can use a low-dimensional information to store, calculate, and transmit the information of the events that have occurred. These state attributes for different events will ultimately affect the dynamic level regulation process of the intelligent agent when facing emergencies.

[0067] 1. In information source situation sharing, it is necessary to solve the problem of information source comparison from different sources and the problem of coding coordination between different information sources.

[0068] Using R i，m represents the virtual information source marked by individual i in the information map of individual i, and m is its number;

[0069] use It represents the virtual information source that is marked by other individuals and transmitted to i, and m is its number.

[0070] R i，m and They share the same numbering system, and the “*” simply represents fuzzy information that is input by other individuals.

[0071] When individual i and individual j meet and they can realize local networking for information transmission, agent i will transfer the information R of information source m to i，m Transmit and share to j. If j has not observed m before, the information source will be marked as n is the number that marks the fuzzy information source in the information map of agent j.

[0072] In order to ensure the coordination requirements between intelligent agents, the marking number m of the target by observer i should be passed to receiver j, so that the local cluster can process the same object number when gathering and attacking.

[0073] Since agents should transfer all information together when exchanging information map data in local networking, the information exchanged between j and k is essentially all information. i，m and They share the same numbering system, so whether they are discovered by themselves or transmitted by other agents, the number of the target will continue as the number of tags increases.i，m The information in should include the time when the target was first discovered, which is used as sorting information.

[0074] like Figure 2 As shown, assume that individuals i and j are at t 1 and t 2 Two information sources are observed at time t and the message is transmitted to individual k, but there is no network between i and j. 1 At time i, the information R about the information source i，1 Passed to k, marked as In t 2 At this moment, j and k come into contact and form a network. j collects the R j，1 Passed to k, it is possible that number 1 has already been used. The time stamp in the information is t 1 , R j，1 The time stamp in the information is t 2 , when j and k exchange information, the information transmitted by j should be placed at the end, so it is marked as The message sent by k should be placed in the R of j's information map list. j，1 In front of, so R j，2 =R j，1 (All numbers after 2 are automatically increased by one). This completes the unification of numbering.

[0075] like Figure 3 As shown in Figure 2, if individuals i and j discover the same unit, the process of transmitting information to k requires checking the input information. Represents the location information of the information source. When the two information sources are considered to be the same, the sum can be taken to get the mean, where r vague is the blur radius. Of course, in addition to integrating information about static targets based on position information, we can also integrate information about moving targets based on speed information. In addition, the newly input virtual information source It also needs all the information of its own mark R j，m For comparison, if The input information is ignored, and only the information source observed by itself is retained and recorded. Through the above operation, the event source in the active event can be marked on the map information of the intelligent body in the form of a virtual information source and transmitted and shared, as if the information of these event sources is uniformly distributed by the background.

[0076] 2 Corresponding to the occurrence of an event, agent i will self-stimulate the corresponding state value Where k is the event type, and corresponding to all event types, the change of the state representation value of individual i over time can be integrated as follows:

[0077]

[0078] The three items on the right correspond to the self-excitation of the state value caused by the occurrence of the event, the conduction of the signal source state, and the decay of the state over time, where α, β, and γ are the mixed modulation coefficients, and δ(t event -t) represents the occurrence process of the event at time t. The conduction process of the signal source state is

[0079]

[0080] As mentioned above, the transmission process of state value is not limited to the transmission between our clusters. i represents the set of all information sources that can interact with individual i, and G is the conduction modulation function. The conduction process represented by T is a more important intercommunication and interaction mechanism under communication constraints. It includes the comprehensive impact of all information sources that an individual can detect on itself. These information sources include both the state representation values ​​of the neighbors around the individual and other identified information sources.

[0081] The individual state value S that is affected and changed according to the event type k k In the propagation interaction, it will be affected by other information sources about the state type, that is, the state propagation process will be affected by the numerous information sources R j Extract And other information to interact.

[0082] Different event types have different roles played by various parts in the differential equation of a certain state change of an individual. According to the task characteristics of large-scale intelligent agent clusters, the state attributes caused by the event types can be roughly divided into four types: "fear", "excitement", "desire" and "disgust". As an extensible algorithm, more state attributes can be set according to more specific task requirements in the application.

[0083] After the intelligent agent discovers a valuable target in the reconnaissance mission, the target will be marked in the information map of individual i, and the virtual information source will in principle attract individual i, and its corresponding "desire state" will be triggered. This process actually finds that the target event has been stimulating the attribute state value during the duration. In the absence of attribute state propagation,

[0084]

[0085] If the desire state of individual i is balanced and stable under continuous stimulation, then Through parameter design, we can give individual i a continuous desire state after discovering a valuable target. Obviously, it is too general to describe an individual's desire for a high-value target only with the representation value of the desire state.

[0086] The difference between active events and passive events is that each active event needs to generate a corresponding state representation based on the event source, that is, for the discovered high-value target and its virtual information source R i，m , the desire state of individual i should be It is used to distinguish the desire state for other high-value goals. To determine which task an individual is going to perform and which goal he or she is approaching, the order of tasks needs to be set by arranging the desire state for each goal.

[0087] In passive events, such as the fear state S triggered by harm to the individual and other neighbors around him f , mainly depends on the transmission process of state information. When individual j detects being attacked, its passive fear state will be stimulated. This state is like a frightened individual in a group of animals spreading fear through violent movements and shouting, which will increase the fear state of the surrounding neighbors. In this event type, the propagation function can be set to

[0088]

[0089] In order to ensure the numerical convergence of state attributes during propagation, the above process represents the fear state S f It can only spread from high to low.

[0090] Similarly, the excited state S e It is also a state attribute that responds to passive events. Its activation process is that if the current task execution phase is not hindered, the state attribute will be activated. For example, if it is not attacked, continues to track high-value targets, conducts fire strikes on the enemy, and explores new areas, all situations that can benefit the current task will continue to activate S. e .

[0091] Disgust state S h Similar to the desire state, its activation process is based on the state of the enemy individuals that have been discovered to track and attack itself, so it is also a response to active events involving specific units. At the same time, because the disgust state is accompanied by attacks and obstruction of mission execution, its fear state and excitement state will often change accordingly.

[0092] 3 The hybrid modulation control will be triggered according to the four corresponding events involved in the current simulation. Figure 4 The state properties shown are mutually stimulating and inhibiting.

[0093] The desire state corresponds to the agent's preference for target selection when discovering a high-value target to select a specific task to perform. The desire state will stimulate the excitement state at the same time.

[0094] The disgust state corresponds to the state triggered when the agent finds and can identify the opponent unit that is interfering with itself. Unlike the desire state, which has a specific preference, the disgust state responds to the disgust state triggered by the interfering unit on the agent itself at the dynamic execution level. If the agent can find the source of the attack, the attack event will trigger both the disgust state and the fear state.

[0095] The excited state corresponds to whether the agent can successfully perform the task. If there is no other interference such as attack, tracking and expulsion, the state will be continuously excited. In other words, when the fear emotion corresponding to the attack event and the disgust emotion of being disturbed are triggered, the excitation process of this state will be suppressed and stopped. At the same time, the unit deeply trapped in "fear" will pay less attention to the excited state of the surrounding neighbors, that is, the fear state will inhibit the transmission of the excited state between agents.

[0096] Similar to the excited state triggered by passive events, the fear state will be triggered as long as the agent is hit. At the same time, the fear state will spread rapidly between agents. The agent can judge the specific area currently hit by the opponent through the distribution of the size of the fear state. When an individual is in an excited state, for example, when it is close to a high-value target, the excited state will promote the decay of the fear state, that is, excitement will accelerate the disappearance of the fear state.

[0097] Scenario settings for simulation tasks:

[0098] This algorithm will simulate the reconnaissance and strike mission of a single high-value target as an example. Because the perception collaboration algorithm is a patent at the algorithm level, the simulation will abstract and simplify the mission assumptions and scenarios to a certain extent, only extracting the obvious features involved in the reconnaissance and strike mission for setting, and will not describe too much engineering details.

[0099] The reconnaissance and strike missions for single high-value targets are:

[0100] 1. The coordinates of the high-value target overlap with those of the defensive firepower and are placed together in the middle of the demonstration area. Their coordinates are initially hidden from unmanned group entry.

[0101] 2 The deployment of the intelligent group is below the demonstration area, and its initial target is above the demonstration area. That is, when the action starts, the system will virtualize a specific coordinate target above the demonstration area. This virtual target will attract all clusters, but as high-value targets are discovered, the virtual target coordinates will be replaced. The intelligent body that completes the initial deployment will obtain full health value (in the simulation, its survival status is represented by HP).

[0102] 3 The behavior mechanism of defensive firepower is: its actual damage per second will decrease with distance, but the splash range will increase, and its damage amount will decrease in the form of

[0103] D * (Δθ, d) = D(1-d * / r d ) 1 / 2 e -Δθ / 2σ (5)

[0104] Where Δθ is the angle between the position vector angle of the agent and the defensive firepower to the agent and the current direction angle of the defensive firepower; d * is the distance between the defense firepower and the agent, r d is the maximum attack distance of defensive firepower, when the distance d * More than r d , defensive firepower cannot search for the intelligent agent and cannot cause damage. σ is the attenuation coefficient of the damage on the fan angle, and D is the default damage per second of the defensive firepower. The defensive firepower will track and lock the intelligent agent units that enter the attack range. The locking mechanism is to predict the maximum threat unit in the distance and the minimum angle rotation amount through the dual weight setting to achieve the goal of striking the closest and most threatening unit in the future with the least possible rotation direction. The maximum angular velocity of the defensive firepower is ω.

[0105] 4. Figure 5 In the simulation scene screenshot shown, the three circles from the inside to the outside are the attack range circles r of the intelligent agent. a , the agent's reconnaissance circle s , the attack range of defensive firepower d Only when the agent enters the reconnaissance circle can the high-value target be discovered, marked, and its information transmitted. Only when the agent enters the strike circle can the agent cause damage to the valuable target.

[0106] D agent =1·ln(e·N rs ) (6)

[0107] The damage capability of the agent is normalized to 1. The more agents that detect valuable targets at the same time, the stronger their locking and attack capabilities will be (often reflected in the accuracy of the attack). Therefore, the damage will be based on the number of surviving agents N in the reconnaissance circle. rs However, the marginal increase in damage from the increase in quantity decays rapidly.

[0108] 5 The entire simulation process is a round from the deployment of the intelligent agent cluster to the death of all agents. In addition to being regulated by the emotional mechanism, the basic dynamic operation logic of the intelligent agent is a self-propelled particle model with a collision volume. After the simulation is completed, the entire execution process will be statistically analyzed in multiple systems.

[0109] The whole simulation process can be explained from the perspective of quantitative indicators as follows: Figure 6 As shown in the figure, the intelligent agent cluster is constantly losing in the process of approaching the high-value target (the remaining life HP in the figure has been scaled and can be expressed as the survival rate). After the intelligent agents continuously enter the attack circle, the ratio of individuals in the circle to the whole continues to increase, while inflicting damage to the high-value target.

[0110] Simulation process:

[0111] 1 After the initial deployment, the agent cluster will be assigned to a virtual destination point. At this time, all agents will be stimulated to desire the destination point. Since this information is directly given by the system, the information characteristics of all individuals for the virtual destination are similar and can be marked as R. i，1 , in the simulation, the desire emotion will drive the individual to approach the unit with the highest desire priority. This attraction is expressed as:

[0112]

[0113] c a is the intensity, r a is the attack range of the agent, is the unit direction vector from the agent to the goal.

[0114] 2. As the agent moves forward, it gradually enters the attack range of the defensive firepower. The group of agents will begin to be damaged, but the source and location of the damage cannot be determined at this time. At this time, the agent will trigger a fear state and spread it to the surrounding neighbors.

[0115]

[0116] The triggering of the fear value is related to the amount of damage the unit receives. If the individual is continuously attacked over a period of time, the state will continue to be triggered.

[0117] The fear-driven individuals will gain a virtual volume expansion in other individuals, which means that the fear unit and the space around it are in danger, so other agents will enlarge the volume of the attacked unit based on the collision avoidance mechanism to avoid being attacked. The dynamic expression of avoidance is

[0118]

[0119] where c d is the intensity of action, the projection here It is the projection between the unit vector in the vertical direction from the individual to the target direction and the unit vector of the relative position from individual j to individual i.

[0120] In the simulation, when the agent cluster cannot know the coordinates of the opponent's defensive firepower, nor does it know the target of concentrated fire and the splash range, the agent can quickly dodge by sensing the fear state of its neighbors. This shows that the single-dimensional information of the fear state can better restore the situation on the battlefield.

[0121] From the density distribution statistics Figure 6 It can be seen that when the intelligent agent cluster gradually enters the attack range of the defensive firepower, the intelligent agent cluster spontaneously spreads rapidly, and gradually concentrates on both sides of the value target in the process of continuously approaching the value target, and its density value on both sides increases significantly. On the one hand, this formation diffusion process can cope with the excessive damage caused by the splash attack of the defensive firepower; on the other hand, the process of the intelligent agents concentrating on both sides and continuously advancing will cause the defensive firepower to continuously turn around on both sides, wasting its attack capability to the greatest extent. This formation with two wings restraining each other provides more survivability for the intelligent agent cluster. This response form that is completely self-organized by the intelligent agent cluster without any preset is the intelligent collaborative self-organizing control behavior emphasized by this algorithm.

[0122] 3 If the agent is not harmed, its excited state will continue to be excited. If the transmission between individuals is not considered,

[0123]

[0124] make γ e =1, then there is a stable state Then the decay coefficient of the fear state is:

[0125]

[0126] As the agent gets closer to the target, its excited state will suppress its fear state more, causing the group of agents to autonomously initiate encirclement behavior when approaching the target.

[0127] Such as velocity projection distribution statistics Figure 8 As can be seen from the figure, when the agent approaches a high-value target, the fear decay coefficient decreases as d decreases. It will rise rapidly, and at the same time suppress the individual's fear state, so that the individual's speed direction is more adjusted to the value target at this time, showing that when it is about to hit the target, the agent can give up the evasive behavior and choose the "rush" behavior, which is an effective strategy conversion. When the evolution time is 500, compared Figure 6 It can be found that at this time all the remaining surviving agents have entered the strike circle, so their speed direction can be adjusted at will.

[0128] 4 When the agent moves forward and finds a high-value target, the target will be marked on its information map by the agent, and the index R i，2 Describe it, and the information will be spread to the surrounding neighbors, and the units receiving the information will be marked on their information map But when they detect the target themselves, they will mark the target they have discovered. j，2 and Merge and organize into R j，2 After local propagation, the cluster will quickly switch its overall target to the currently detected high-value unit.

[0129] 5 When the agent detects the coordinates of the opponent's defensive firepower, it will stimulate disgust towards the defensive unit. Without considering the propagation process, Similar to the Desire state, when this state attribute is continuously stimulated, its stable value is S h =α h / γ h , the difference is that individuals can stimulate disgust towards multiple interference units at the same time and act together. The process of stimulating the disgust state is related to the threat posed by the interference unit to the intelligent agent. The most intuitive relationship is the spatial distance. is the aversion escape distance parameter of the jammer source l. The closer the distance between the jammer and the agent, the greater the threat. Accordingly, a self-excitation effect can be established:

[0130] where d * is the distance between the agent and the interferer.

[0131] The control model of the aversion state on the dynamics of the intelligent agent is:

[0132]

[0133] l is the interference unit, G i is the set of interference units marked in the information map of i that can have a negative impact on itself, c h is the intensity of action, the projection here is the unit vector in the direction of the relative position of the interference unit to the agent. When the hostile interference source, such as defensive artillery, coincides with the coordinates of a high-value target, the repulsive force caused by disgust will not work.

[0134] 6 After entering the attack circle, the agent will cause damage to the valuable target, and a single simulation will stop when all the agents in the team are killed.

[0135] From the simulation results, Figure 6 , 7 , 8, it can be seen that the swarm intelligence collaborative algorithm can effectively realize the information dimensionality reduction processing, collaborative transmission and restoration of complex situations during information processing of intelligent agents for complex battlefield situations; on this basis, the cluster can realize the self-organizing intelligent collaborative combat behavior in the absence of all information within the framework of distributed computing by regulating the dynamic system influenced by the state attributes of the emotion-like mechanism; through the mutual adjustment of different state attributes, the intelligent agent cluster can also realize the switching of tactical strategies to a certain extent.

Claims

1. A distributed computing method for collaborative situational awareness of intelligent agent clusters, characterized in that: It includes three contents: information source status sharing, including information source confirmation, sharing and information transmission selection; event status sharing, the state attributes that stimulate propagation based on events and their types; the hybrid modulation process of mutual influence between states; To realize situational awareness and collaboration, the intelligent agent cluster must first improve the annotation of data information, which requires the establishment of an information annotation and exchange system based on the local networking of intelligent agents. Based on the information exchanged between intelligent agents and complex environmental information, an abstract state representation system that can reflect real-time events is constructed, and combined with the information exchange system of the local networking, the intelligent agents can realize the mutual influence of individual state information based on event situations. By introducing other characteristic state values ​​in the process of excitation, propagation and decay of abstract state representation values ​​and establishing a set of dynamic models regulated by characteristic states, we can obtain an intelligent group adaptive collaborative control method with interpretability and complex performance in battlefield environments. Information source situation sharing includes the following steps: Step 1: Establish the agent’s own local database, including individual information maps for labeling and coordination, and label the information source of newly discovered targets for data transfer; Step 2: The agents use the communication function to conduct local networking and data transmission, and number the information in different forms according to the different sources of information; Step 3: Coordinate and unify the order of information source numbers within the agent cluster by introducing the first marking time of the information source In order to ensure the coordination requirements between agents, the target mark number m of observer i should be transmitted to receiver j, so that the local cluster will process the same number of objects when gathering and attacking, because R i，m and They share the same numbering system, so whether they are discovered by themselves or transmitted by other agents, the number of the target will continue as the number of tags increases. Assume that individuals i and j observe information sources a and b at time t1 and t2 respectively, and pass the message to individual k, but there is no network between i and j. At time t1, i sends the information R about information source a to individual k. i，m Passed to k, marked as At time t2, j and k come into contact and form a network. j collects the R j，m Passed to k, at this time, the number m has been used, because The time mark in the information is t1, R j，m The time in the message is marked as t2. When j and k exchange information, the information transmitted by j should be ranked later, so it is marked as The message sent by k should be placed in the R of j's information map list. j，m In front of, so R j，m+1 =R j，m , all numbers after m are automatically increased by one, Step 4: Solve the problem of repeated labeling of a single source by merging similar information; When individuals i and j discover the same unit, the process of transmitting information to k requires checking the input information. Represents the location information of the information source. When the two information sources are considered to be the same, the sum is taken to get the average, where r vague To fuzzy the radius, the static target is integrated according to the position information, and the moving target is integrated according to the speed information. In addition, the newly input virtual information source It also needs all the information of its own mark R j，m For comparison, if The input information is ignored, and only the information source observed by itself is retained and recorded. The event source in the active event is marked on the map information of the intelligent body in the form of a virtual information source and transmitted and shared; Incident situation sharing includes the following steps: Step 1: Intelligent individuals pass through specific events and the corresponding characteristic state values ​​based on them To record the local situation information of the battlefield, where k is the event type, corresponding to all event types, establish any feature state Differential equation evolving over time: The three items on the right correspond to the self-excitation of the state value caused by the occurrence of the event, the conduction of the signal source state, and the decay of the state over time, where α, β, and γ are the mixed modulation coefficients, and δ(t event -t) represents the occurrence process of the event at time t; Step 2: According to different event types, establish a state information transmission mode based on feature states; Step 3: Distinguish the events into active events and passive events according to the event type; Step 4: For source events, the state attributes need to further mark the state source; Hybrid modulation control includes the following steps: Step 1: When individual i finds the value target l, his desire state It will inspire; This property will adjust its coefficient α by the distance d desire , α desire =e -d , by comparing the state attributes of each target Then, multiple targets can be sorted by their desired states and target m can be finally selected; Step 2: The individual is attracted by the target unit. The desire state drives the individual to approach the unit with the highest desire priority. The attraction is expressed as: c a is the intensity, r a is the attack range of the agent, is the unit direction vector from the agent to the target; Step 3: When the agent is attacked, its fear state S f It will inspire; The excitement state will adjust the value of the fear value to The mixed modulation process of excitation, propagation and decay is: The triggering of the fear value is related to the amount of damage D* / D received by the unit. If the individual is continuously attacked over a continuous period of time, the state will continue to be triggered; Step 4: Fear-driven individuals gain virtual volume expansion in the judgments of other individuals; The damage event means that the fear unit and its surrounding space are in danger, so other agents will enlarge the volume of the units that are actually attacked based on the collision avoidance mechanism to avoid being attacked. The dynamic expression of avoidance is c d is intensity, projection is the unit vector in the vertical direction from the individual to the target The unit vector of the relative position of individual j to individual i The projection between Step 5: When the agent is not affected and can successfully perform the task, its excited state will continue to be excited; Step 6: When the agent is tracked, attacked, or disturbed by a unit that has been discovered by the other party, the agent will generate disgust toward the disturbing unit. The process of triggering the disgust state is related to the threat posed by the interference unit to the agent. The most intuitive relationship is the spatial distance. is the aversion escape distance constant of the interferer source l. The closer the interference unit is to the agent, the greater the threat. The corresponding self-excitation effect is where d * is the distance between the agent and the interferer; Step 7: The expression of the aversion state for the agent dynamics is: G i is the set of interference units marked in the information map of i that can have a negative impact on itself, c h is the action intensity, projection is the unit vector of the relative position of the interference unit to the agent; Step 8: In other cases, the agent is driven by a self-propelled particle dynamics model with physical collisions.

2. The distributed computing method for collaborative situational awareness of a cluster of intelligent agents according to claim 1, characterized in that: Step 1 in information sharing: Use R i Represents the basic information of individual i, including the location of the individual speed And other state representation quantities corresponding to event types k represents the event and state type. The agent cluster has marked all individuals in the system during the initial deployment, so the agents can recognize each other and determine their unique numbers. When the agent discovers the opponent's high-value target or the opponent's interception unit during operation, the information source of the event, that is, the discovered target, will be marked as a node of state information in the agent's information map. For agent i, its reconnaissance mark is R i，m , which contains all the information after individual i detects and virtually evaluates target m, and is used represents the virtual information source that is marked by other individuals and transmitted to i, m is its number, R i，m and They share the same numbering system, and the "*" simply represents fuzzy information that is input by other individuals.

3. The distributed computing method for collaborative situational awareness of a cluster of intelligent agents according to claim 1, characterized in that: Step 2 in information source situation sharing: When individuals i and j meet and can realize local networking for information transmission, agent i will share the information R of information source m. i，m Transmit and share to j. If j has not observed m before, the information source will be marked as n is the number marked by information source m in the information map of agent j.

4. The distributed computing method for collaborative situational awareness of a cluster of intelligent agents according to claim 1, characterized in that: Step 2 in event situation sharing: The transmission process of the signal source state is C i represents the set of all information sources that can interact with individual i, G is the conduction modulation function, and the individual state value S that is affected and changed according to event type k k In the propagation interaction, it will be affected by other information sources about event type k, that is, the state propagation process will be affected by the numerous information sources R j Extract Information interaction.

5. The distributed computing method for collaborative situational awareness of a cluster of intelligent agents according to claim 1, characterized in that: Step 3 in event situation sharing: Different event types have different effects on the various parts of the differential equation of a certain state change of an individual. According to the task characteristics of large-scale intelligent agent clusters, the state attributes caused by the event types are divided into passive events such as unknown damage events and the corresponding "fear" state, unimpeded task execution and the corresponding "excitement" state, and specific target detection and approach behavior and the corresponding "desire" state, and specific interception or expulsion events and the "disgust" state stimulated by the interfering individual.

6. The distributed computing method for collaborative situational awareness of a cluster of intelligent agents according to claim 1, characterized in that: Step 4 in event situation sharing: Active events need to generate corresponding state representations based on the event source, that is, for the discovered high-value targets and their virtual information sources R i，m , the desire state of individual i should be It is used to distinguish its desire state for other high-value targets. Through the process of stimulating the desire state of the cluster by different information source targets and the propagation between agents, it is possible to determine the information source with the greatest attraction for a single agent as the current task target. The disgust state S h Similar to the Desire state, its activation process is based on the state of the enemy individuals that have been discovered to track and attack itself, so it is also a response to active events involving specific units.

7. The distributed computing method for collaborative situational awareness of a cluster of intelligent agents according to claim 1, characterized in that: In step 5 of the hybrid modulation control, if the transmission between individuals is not considered, then When the state is stable, S e =α e / γ e , considering that the closer the individual is to the goal, the higher the excitement level will be, so we set Let γ e =1, then When an individual is in a "panic" state with a high fear value, the individual cannot obtain normal excitement from the surrounding neighbors, that is, fear will inhibit the interaction of excitement attributes between intelligent agents.

Images