Multi-agent role autonomous negotiation and scheduling method and system based on asynchronous communication

Abstract

The application discloses a multi-agent role autonomous negotiation and scheduling method and system based on asynchronous communication, relates to the technical field of multi-agent system cooperative control and distributed decision-making, and comprises the following steps: constructing an asynchronous state synchronization model; each agent locally maintains a perception state stream and deploys a lightweight time sequence prediction model to perform real-time prediction on the perception state stream; based on a role handover event or a cooperative request event generated in advance, an asynchronous negotiation and scheduling process is triggered; asynchronous negotiation: a streaming negotiation mechanism is established to complete asynchronous dynamic allocation of roles online; scheduling process: distributed arbitration is performed according to a logic vector clock or a hash value, and a conflict-free replication data type state is rolled back; soft mounting and hard switching are adopted to realize seamless handover of roles; and the method aims to solve technical problems such as poor scalability and easy deadlock of multi-agent cooperation in a dynamic and decentralized environment due to dependence on a global clock and centralized scheduling.

Description

Technical Field

[0001] This invention relates to the field of collaborative control and distributed decision-making technology for multi-agent systems, and particularly to a method and system for autonomous negotiation and scheduling of multi-agent roles based on asynchronous communication. Background Technology

[0002] With the rapid development of artificial intelligence and unmanned systems technology, multi-agent systems (MAS) are playing an increasingly important role in complex scenarios such as drone swarm formation, large-scale logistics and warehousing scheduling, deep-sea / deep-space exploration, and disaster emergency search and rescue. In these decentralized application scenarios, how to efficiently and dynamically allocate roles and schedule tasks is the key to determining the overall survivability and mission efficiency of the swarm.

[0003] However, existing multi-agent cooperation and task scheduling mechanisms still face significant technical bottlenecks when dealing with complex physical environments characterized by high dynamics and weak communication. These bottlenecks are manifested in the following aspects: First, traditional collaborative control heavily relies on centralized node coordination or strict global physical clock synchronization (such as NTP / PTP protocols). In real-world environments such as the field, underground, or with strong electromagnetic interference, communication links often suffer from high latency, high packet loss rates, or even partial network outages. In such cases, systems relying on a global clock will frequently experience misjudgments due to timestamp misalignment, while centralized scheduling systems face serious single-point-of-failure risks, making it difficult to meet the robustness and boundless scalability requirements of large-scale clusters.

[0004] Secondly, in existing decentralized task negotiation mechanisms (such as the traditional Contract Net Protocol), task initiators typically employ a synchronous blocking decision-making approach. This means that after issuing a tender, they must wait indefinitely for responses from all potential nodes before performing a global comparison to find the best option. This rigid synchronous waiting mechanism is highly susceptible to indefinite suspension (i.e., deadlock) in asynchronous networks due to network latency or disconnections at some edge nodes. Furthermore, existing task triggering mechanisms are mostly based on delayed responses with fixed thresholds (e.g., initiating role handover only when device battery is nearly depleted). This passive triggering cannot compensate for the inherent time delays of asynchronous communication, frequently leading to task handover failures or cluster paralysis.

[0005] Finally, in completely decentralized asynchronous networks, state consistency and concurrent conflicts are another pressing problem to be solved. When multiple agents, based on partially incomplete information, simultaneously decide to compete for the same key role within a similar timeframe, traditional distributed consensus protocols (such as Paxos or Raft) require multiple rounds of large-scale network communication to acquire distributed locks. This incurs significant bandwidth overhead and is prone to timeouts and blocking in weak communication environments. Furthermore, once communication breaks down, causing some agents to become isolated network islands, the system often lacks deterministic conflict resolution rules and seamless hardware / software handover mechanisms, making it highly susceptible to system logic collapse or physical collisions.

[0006] Therefore, there is an urgent need to propose a multi-agent role scheduling method and system that does not rely on a global physical clock, has lock-free state synchronization capability, and can achieve asynchronous streaming negotiation through predictive event-driven methods, in order to solve the defects existing in the above-mentioned technologies. Summary of the Invention

[0007] Based on the technical problems existing in the background technology, this invention proposes a method and system for autonomous negotiation and scheduling of multi-agent roles based on asynchronous communication, aiming to solve the technical pain points of poor scalability and easy deadlock caused by multi-agent collaboration in dynamic and decentralized environments due to reliance on global clock and centralized scheduling.

[0008] The multi-agent role autonomous negotiation and scheduling method based on asynchronous communication proposed in this invention includes: Construct an asynchronous state synchronization model based on conflict-free copy data type and rumor protocol to achieve lock-free consistent update of local state of multiple agents in a decentralized environment; Each agent locally maintains a perception state stream based on a sliding time window and deploys a lightweight time-series prediction model to make real-time predictions. When the prediction result exceeds the dynamic adaptive threshold, a role handover event or a collaboration request event is generated in advance to trigger the asynchronous negotiation and scheduling process. Asynchronous negotiation: Establish a streaming negotiation mechanism based on the asynchronous continuous contract network protocol, and combine time decay utility function and rolling clearing strategy to complete the asynchronous dynamic allocation of roles online; Scheduling process: Distributed arbitration is performed based on logical vector clock or hash value, and the state of conflict-free replicated data types is rolled back; soft mounting and hard switching are used to achieve seamless role handover; when isolated, survival tasks are downgraded, and the state is merged and re-entering the network after the network is restored via rumor protocol.

[0009] Furthermore, the construction of the asynchronous state synchronization model based on conflict-free copy data type and rumor protocol includes: Logical vector clocks are used to assign partial-order timestamps to local events of the agent; The role occupancy state of multiple agents and the task scheduling queue are modeled as a conflict-free copy data type, which supports agents to update state data locally without blocking. Based on the rumor protocol, state increments are asynchronously propagated within the local communication radius of the agent, and historical conflicts are resolved through mathematical merging operations to achieve eventual consistency of the network state.

[0010] Furthermore, the update rule for the logical vector clock is specifically as follows: When the intelligent agent When a local event occurs, increment the local counter. When the intelligent agent When sending a message, the current logical vector clock is attached to the message message; When the intelligent agent Received from the intelligent agent When a message is received, the local clock is updated by taking the maximum value of each element, and the local counter is incremented.

[0011] Furthermore, the perception state stream includes its own resource consumption rate, target movement trajectory, and environmental threat level.

[0012] Furthermore, using the perceived state flow as input, the lightweight time-series prediction model extracts time-series evolution features to predict in real time the probability of task failure or the time of resource exhaustion within a set future time window.

[0013] Furthermore, in the asynchronous negotiation, a streaming negotiation mechanism based on the asynchronous continuous contract network protocol is established, including: The task initiator broadcasts task characteristic information, including a logical vector clock and a negotiated deadline, to the local network using a publish / subscribe model. The bidding agent calculates the utility value of the task by combining local execution capabilities with a time decay factor; The task initiator adopts a rolling clearing decision mechanism that does not wait for all nodes to bid. When the highest efficiency value received within the set rolling time window exceeds the satisfaction threshold, a confirmation message is sent asynchronously to achieve streaming task matching.

[0014] Furthermore, the distributed arbitration based on logical vector clocks or hash values ​​includes: When multiple intelligent agent nodes concurrently compete for the same role, the causal order is determined based on the partial order relationship between two logical vector clocks; If there is a sequential relationship, the operation that occurs later is considered valid. If there is no order, the hash values ​​of the agent identifiers are sorted in descending order, and the hash value at the top wins. The loser in the arbitration automatically reverts to the state of a conflict-free copy data type.

[0015] Furthermore, the scheduling process employs soft mounting and hard switching to achieve seamless role handover, including: The original executor retains control of the mission; The successor silently establishes a perceptual connection with the task environment in the background and asynchronously receives the local environment context packaged and sent by the old role executor. Control is transferred instantly after the asynchronous handshake is confirmed.

[0016] Furthermore, the isolation and degradation of the survival task includes: when the duration of the communication interruption between the agent and other agents reaches a preset timeout, the agent is determined to be in a communication isolation state; the agent automatically terminates the waiting asynchronous negotiation, and determines and executes an independent survival task according to the remaining resource status and environmental threat status obtained locally, based on preset degradation rules, such as silent return, in-situ concealment, or single-machine local patrol.

[0017] A computer system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the method described above.

[0018] The advantages of the multi-agent role autonomous negotiation and scheduling method and system based on asynchronous communication provided by this invention are as follows: It breaks the limitation of the global physical clock and realizes a lock-free architecture with extremely high concurrency; It abandons the traditional global physical clock and innovatively combines logical vector clock and conflict-free copy data type (CRDT), enabling agents to update role states locally with zero blocking; This lock-free mechanism guaranteed by the mathematical underlying layer completely eliminates the communication blocking and deadlock risks brought by traditional distributed locks, and gives the multi-agent cluster extremely strong scalability in harsh network environments; Transforming delayed response into proactive scheduling greatly enhances the system's extreme anti-latency capability: This embodiment differs from traditional fixed threshold alarms by introducing a lightweight time-series prediction model at the edge. Through real-time reasoning of the sliding state flow, the system can predict nodes that will run out of resources or fail in advance and proactively trigger negotiation events in advance. The time difference gained through pre-scheduling perfectly offsets the inherent data latency of asynchronous networks. Eliminating the bottleneck of synchronous waiting, this embodiment achieves millisecond-level streaming matching in a decentralized environment: It proposes an Asynchronous Continuous Contract Network (A-CNP) protocol, where task initiators do not need to wait for bid responses from all nodes, but instead dynamically clear out participants based on a time decay utility function and a satisfaction threshold. This not only avoids negotiation suspension due to the disconnection of some nodes, but also naturally makes cluster behavior more inclined towards fast and localized collaboration. It possesses strong conflict self-healing and island reorganization capabilities: When multiple agents experience role preemption conflicts in an asynchronous state, this embodiment can instantly resolve the conflict and revert to the previous state locally through deterministic distributed arbitration rules. Combining a soft-mount-hard switching mechanism and preset degradation rules in communication island states, the system not only ensures the continuity of role handover but also enables nodes to perform independent survival tasks after becoming communication islands and automatically rejoin the cooperative network when the network recovers. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will now be described in detail through specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0021] like Figure 1 As shown, the multi-agent role autonomous negotiation and scheduling method based on asynchronous communication proposed in this invention includes: S1: Construct an asynchronous state synchronization model based on conflict-free copy data type and rumor protocol to achieve lock-free consistent update of local state of multiple agents in a decentralized environment. S2: Each agent locally maintains a perception state flow based on a sliding time window and deploys a lightweight time-series prediction model to make real-time predictions. When the prediction result exceeds the dynamic adaptive threshold, a role handover event or a collaboration request event is generated in advance to trigger the asynchronous negotiation and scheduling process. S3: Asynchronous negotiation: Establish a streaming negotiation mechanism based on the asynchronous continuous contract network protocol, and combine time decay utility function and rolling clearing strategy to complete the asynchronous dynamic allocation of roles online; S4: Scheduling process: Distributed arbitration is performed based on logical vector clock or hash value and the state of conflict-free replicated data types is rolled back; soft mounting and hard switching are used to achieve seamless role handover; when isolated, the state is downgraded to perform survival tasks, and after the network is restored, the state is merged and re-enter the network through the rumor protocol.

[0022] This embodiment provides a multi-agent role scheduling method and system that does not rely on a global physical clock, has lock-free state synchronization capability, and can achieve asynchronous streaming negotiation through predictive event-driven methods. It aims to solve the technical pain points of multi-agent collaboration in dynamic and decentralized environments, such as poor scalability and susceptibility to deadlocks due to reliance on a global clock and centralized scheduling.

[0023] Breaking the limitations of the global physical clock, it achieves a lock-free architecture with extremely high concurrency: abandoning the traditional global physical clock, it innovatively combines logical vector clocks with conflict-free copy data types (CRDT), enabling agents to update their role states locally with zero blocking; this lock-free mechanism, guaranteed by the underlying mathematics, completely eliminates the communication blocking and deadlock risks brought by traditional distributed locks, and gives multi-agent clusters extremely strong scalability in harsh network environments.

[0024] Transforming delayed response into proactive scheduling greatly enhances the system's ability to withstand extreme latency: This embodiment differs from traditional fixed threshold alarms by introducing a lightweight time-series prediction model at the edge. Through real-time reasoning of the sliding state flow, the system can predict nodes that will run out of resources or fail in the task in advance and proactively trigger negotiation events in advance. The time difference gained through pre-scheduling perfectly offsets the inherent data latency of asynchronous networks.

[0025] Eliminating the bottleneck of synchronous waiting and realizing streaming matching in a decentralized environment: This embodiment proposes an Asynchronous Continuous Contract Network (A-CNP) protocol. Task initiators do not need to wait for bid responses from all nodes; instead, dynamic rolling clearing is performed based on a time decay utility function and a satisfaction threshold. In disaster emergency search and rescue scenarios, the tasks include area search tasks, target tracking tasks, and communication relay tasks, corresponding to area search roles, target tracking roles, and communication relay roles, respectively. The task initiator is an agent that initiates new task allocation, role handover, or collaboration requests. This not only avoids negotiation suspension due to the disconnection of some nodes but also makes cluster behavior tend towards rapid, proximity-based collaboration.

[0026] It possesses strong conflict self-healing and island reorganization capabilities: When multiple agents experience role preemption conflicts in an asynchronous state, this embodiment can instantly resolve the conflict and revert to the previous state locally through deterministic distributed arbitration rules. Combining a soft-mount-hard switching mechanism and preset degradation rules in communication island states, the system not only ensures the continuity of role handover but also enables nodes to perform independent survival tasks after becoming communication islands and automatically rejoin the cooperative network when the network recovers.

[0027] It is understood that the intelligent agent in this embodiment can be any physical entity with autonomous decision-making capabilities, such as a drone, robot, or unmanned vehicle. Each intelligent agent includes: a perception module for collecting local state data; a communication module for asynchronous message sending and receiving with neighboring intelligent agents; a storage module for maintaining conflict-free copy data types; and an execution module for executing assigned role tasks. This embodiment is a method for autonomous role negotiation and scheduling based on asynchronous communication in a multi-agent system containing multiple intelligent agents.

[0028] In one embodiment, step S1 includes steps S1.1 to S1.3.

[0029] S1.1: Use a logical vector clock to assign partial-order timestamps to local events of the agent, eliminating the multi-agent system's dependence on the global physical clock; In this embodiment, the system does not rely on any NTP (Network Time Protocol) for physical clock synchronization. Each intelligent agent... Maintain a logical vector clock locally. ,in, The number of known agents within the local network. For intelligent agents Event count.

[0030] Logic Vector Clock The update rules are (a1) to (a3): (a1) When the intelligent agent When a local event occurs (such as discovering a new task or initiating role negotiation), its local counter increments: ; (a2) When the intelligent agent Send messages At that time, the current logical vector clock will be... Attached to the message message, i.e., sent ; (a3) When the intelligent agent Received from the intelligent agent News At that time, update the local clock by taking the maximum value of each element and increment the local counter: ,and ;in, To iterate through the indices of all known agent IDs, For intelligent agents What is known, intelligent agents The total number of events that have occurred (logical time count). For intelligent agents What is known, intelligent agents The total number of events that have occurred (logical time count). In Indicates that the intelligent agent Increment its own event count by 1 to record the newly occurring local event of receiving a message.

[0031] By using this logical vector clock, the system can accurately capture the causal partial order relationship between events without an absolute time reference, thus completely solving the problem of misjudgment of state caused by out-of-order message arrival in asynchronous networks.

[0032] Step S1.2: Model the role occupancy state and task scheduling queue of the multi-agent as a conflict-free copy data type (CRDT) to support agents to update state data locally without blocking.

[0033] To achieve lock-free concurrent writing, this embodiment abstracts the shared role allocation table and task state set among multiple agents into a state-based conflict-free replication data type, specifically employing the LWW-Element-Set (last writer wins element set) data structure. Specifically, the role allocation record corresponding to each task is used as a set element, which includes at least a task identifier, role identifier, executing agent identifier, role state, and the logical vector clock generated in step S1.1. When a role is assigned, transferred, or released, the addition or deletion status of the corresponding role allocation record is written to the local copy. For multiple records with the same task identifier and role identifier, the valid record is determined based on the logical vector clock during state merging; for concurrent records where the logical vector clocks do not have a causal relationship, the valid record is determined based on the hash value of the agent identifier.

[0034] In the LWW-Element-Set data structure, each entry of a role's state (e.g., Task A - Assigned to - Agent B) not only contains the data itself but is also bound to a timestamp of the logical vector clock generated in step S1.1. When an agent decides to take over a role, it does not need to request a lock from the entire network; instead, it directly writes the state change to a local, conflict-free, replicated data type copy. Because the conflict-free replicated data type satisfies the commutative, associative, and idempotent laws, the merging operation of any two concurrent states is mathematically guaranteed to converge to the same result. This design allows agents to... This allows for local non-blocking writes with a time complexity that greatly improves the system's concurrent throughput.

[0035] Step S1.3: Based on the Gossip Protocol, asynchronously propagate the state increment within the local communication radius of the agent, resolve historical conflicts through mathematical merging operations, and achieve eventual consistency of the network state.

[0036] Since the agent only updates the state of a conflict-free copy data type locally, the system introduces a rumor protocol for asynchronous distribution of the state.

[0037] In each fixed heartbeat cycle (without global alignment, each agent executes according to its local timer), the agent randomly selects a target within its communication radius. Each neighboring node packages its latest conflict-free replicated data type state increment (Delta-CRDT) and sends it to its neighbors. Upon receiving the increment, the neighboring node calls the conflict-free replicated data type merge function to merge the new state with its local state without conflict, and uses a logical vector clock to determine the causal order of the data.

[0038] Through this decentralized model, similar to viral transmission, any local negotiation results and state changes will be recorded. Within a network cycle, the consensus spreads exponentially across the entire network, thus achieving eventual consistency in a multi-agent system under asynchronous conditions without the need for a central node to coordinate. This represents the total number of agents in a multi-agent system (i.e., the number of nodes in the network).

[0039] In one embodiment, in a traditional threshold-triggered mechanism, the agent often initiates a collaboration request only when facing resource depletion (such as extremely low battery) or impending task failure. Due to the inherent information transmission delay in asynchronous communication networks, this delayed response can easily lead to insufficient rescue and task failure. To address this, this embodiment introduces an event-driven mechanism in step S2, specifically including the following sub-steps S2.1 to S2.3.

[0040] S2.1: The agent maintains a perception state stream locally based on a sliding time window, the state stream including its own resource consumption rate, target movement trajectory and environmental threat level.

[0041] Specifically, each agent allocates a fixed-length sliding time window in its local memory to continuously record historical operational data of itself and its surrounding environment. This data forms a time-series perception state stream, mainly covering three dimensions: first, the internal state dimension (i.e., its own resource consumption rate), such as the current remaining battery power and processor resource consumption rate; second, the task object dimension (i.e., the target's movement trajectory), such as the movement trajectory and speed changes of the target being tracked or processed; and third, the external environment dimension (i.e., environmental threat level), such as the density of obstacles detected by sensors, the level of environmental threat, or the attenuation of communication signal strength. By maintaining such a dynamically updated historical data stream, the agent can fully grasp the changing trends of the system state, rather than relying solely on the current static snapshot.

[0042] Step S2.2: Deploy a lightweight time-series prediction model at the edge to predict the probability of task failure or the time of resource exhaustion within a set future time window in real time.

[0043] Considering that edge computing resources for agents are typically limited, this embodiment deploys a lightweight time-series prediction model with low computational consumption in the local controller of each agent. This model could be a lightweight neural network with parameter pruning or an improved Kalman filter algorithm (e.g., EKF (Extended Kalman Filter), UKF (Unscented Kalman Filter), and various deep Kalman filters). This prediction model takes the perceived state stream within the sliding time window obtained in step S2.1 as input, extracts temporal evolution features, and infers and outputs the system's operating state over a future look-ahead period. Specific output prediction metrics include: the probability of failure when continuing to execute the existing role's task under the current trend, and the expected resource depletion time of the agent's core resources (such as battery power). This prediction and inference process is completed entirely independently on the agent's local machine, without communication with external nodes, ensuring extremely high real-time performance.

[0044] Step S2.3: When the predicted probability or prediction time exceeds the dynamic adaptive threshold, generate a role handover event or a collaboration request event in advance, transform the delayed response into a forward-looking schedule, and trigger the asynchronous negotiation process.

[0045] The system sets corresponding safety thresholds for the aforementioned prediction metrics. To adapt to the ever-changing dynamic environment, these thresholds are not fixed, rigid values, but can be dynamically and adaptively adjusted based on the complexity of the agent's current environment, network latency, and task priority. Once the task failure probability output by the local prediction model exceeds the currently set safety threshold, or the expected resource exhaustion time is shorter than the set safety buffer time, the agent will determine that a risk is imminent. At this point, the agent will proactively and in advance generate a trigger event locally without waiting for the failure to actually occur. Depending on the type of risk, this trigger event can manifest as a role handover event (i.e., finding an agent with a better state to take over its current role) or a cooperation request event (i.e., calling on other agents to provide supplementary capabilities). Once the role handover event or cooperation request event is generated, it will immediately wake up and trigger the subsequent asynchronous streaming negotiation mechanism.

[0046] Step S2 cleverly utilizes the time difference gained through forward scheduling by introducing a time-dimensional prediction, effectively offsetting the unavoidable data delay in asynchronous network communication.

[0047] In one embodiment, the traditional Contract Network Protocol (CNP) often requires the task initiator to wait for bid responses from all potential participants in the network during the bidding phase before performing a global optimization to determine the final contractor. This rigid synchronous waiting mechanism can easily lead to the initiator getting stuck in endless waiting or even deadlock in asynchronous networks with high latency or packet loss. Therefore, this embodiment proposes step S3, specifically including sub-steps S3.2 to S3.3.

[0048] Step S3.1: The task initiator broadcasts task feature information, carrying a logical vector clock as a timestamp and a negotiated deadline, to the local network in a publish / subscribe mode.

[0049] When the predictive event in step S2 is triggered (i.e., the generated role handover event or collaboration request event), or when a new external task is injected into the system, the corresponding agent transforms into a task initiator. This initiator no longer uses point-to-point polling communication, but instead broadcasts task bidding information (i.e., publishing topics) to neighboring nodes within its communication radius based on a local publish / subscribe network model. This bidding information data packet not only contains the specific characteristics of the task (such as task type, required computing power, target location, etc.), but more importantly, it includes the partial-order timestamp of the initiator's current local event allocation generated in step S1 (i.e., the current value of the logical clock), and a relative negotiation deadline set based on the initiator's local counter. This publishing mechanism, which does not rely on global absolute time, ensures that bidding information can be asynchronously disseminated in the local network with minimal overhead.

[0050] Step S3.2: The bidding agent calculates the utility value of executing the task by combining local execution capability and time decay factor, wherein the time decay factor causes the utility value of the delayed response to decrease dynamically as physical time passes.

[0051] When other idle agents (i.e. subscribers) in the surrounding network asynchronously receive the above bidding information, they become potential bidding agents. The bidding agent first evaluates a basic execution utility value (i.e., the cost-effectiveness or success rate of executing the task) based on its current idle resources, distance from the task target, and its own capability list.

[0052] To adapt to asynchronous communication environments, this embodiment innovatively introduces a time decay factor in the calculation of the basic utility value. Specifically, when preparing to generate a bid response, the bidding agent assesses the communication latency and processing time experienced from the issuance of the bidding information to the current preparation of the response. As time elapses, the time decay factor dynamically reduces the agent's final bid utility value. This design logically creates a natural selection mechanism: agents that are close to the network, have good network quality, and fast processing speed can submit high-scoring bids; while nodes at the network edge with significant communication latency will have their bid scores automatically reduced over time, thus encouraging the entire cluster to favor fast and nearby collaboration.

[0053] Step S3.3: The task initiator adopts a rolling clearing decision mechanism that does not wait for all nodes to bid. When the highest bid utility value received within the set rolling time window exceeds the satisfaction threshold, a confirmation message is sent asynchronously to achieve streaming task matching.

[0054] This step changes the centralized decision-making model of traditional protocols. After broadcasting the bidding information, the task initiator opens a rolling reception window locally and pre-sets a satisfaction threshold (i.e., an acceptable utility score that meets the minimum requirements of the task). Within the rolling window, the initiator will receive bids from various bidding agents one after another, asynchronously.

[0055] The initiator adopts a streaming processing strategy of receiving and evaluating bids simultaneously: as soon as the utility value of a received bid exceeds a set satisfaction threshold, the initiator immediately stops waiting for that task, asynchronously sends a confirmation message to the high-scoring bidder, and formally assigns it the corresponding task role. If, after the entire rolling time window ends, none of the received bids reach the satisfaction threshold, the highest-scoring bidder is selected as the winner, or a new round of downgraded bidding is triggered. This rolling clearing streaming matching mechanism eliminates global waiting deadlocks in asynchronous networks and achieves millisecond-level dynamic role allocation.

[0056] In one embodiment, in a fully asynchronous network, due to message propagation delays, concurrent conflicts can easily occur where two or more agents simultaneously decide to take over the same task role within a similar timeframe. To resolve this issue and ensure the continuity of task execution without relying on a central arbitration node, step S4 is introduced, specifically including sub-steps S4.1 to S4.3.

[0057] S4.1: When multiple agent nodes concurrently compete for the same role in a state conflict, deterministic distributed arbitration is performed based on the partial order relationship of two logical vector clocks or the size of the local hash value. The loser of the arbitration is automatically demoted and rolled back to its local conflict-free replicated data type state.

[0058] When the system synchronizes state data via a rumor protocol, a state conflict occurs if an agent detects multiple parallel write operations to the same role slot in its local conflict-free replicated data type (CRDT). In this case, the agent does not need to initiate additional negotiation communication but directly triggers a deterministic distributed arbitration algorithm locally.

[0059] The primary basis for arbitration is comparing the logical vector clocks associated with the events that caused the conflict. If the two clocks have a clear causal order, the operation that occurred later is directly deemed valid. If the two clocks are concurrent (i.e., they do not have a sequential relationship), a pre-defined rule for resolving the conflict is used. For example, the hash values ​​of the unique identifiers (IDs) of the agents that generated the event are compared and sorted in descending order, with the hash value at the top winning. Based on this network-wide consistent local deterministic rule, all relevant nodes can reach the same arbitration result locally. Subsequently, the agent that lost the arbitration automatically revoked its preemption of the role, proactively downgraded to an idle state, and reverted its local CRDT state to the valid state before the conflict occurred, thus mathematically eliminating the conflict.

[0060] Step S4.2: When handing over roles, a mechanism combining soft mounting and hard switching is adopted. The successor is in an asynchronous synchronous environment context in the background. After the asynchronous handshake is confirmed, the seamless handover of control is completed instantly.

[0061] When roles successfully negotiate and need to hand over tasks, in order to prevent task interruption or loss of control of agent behavior due to state asynchrony at the moment of handover, this method adopts a soft-mount-hard-switching flexible handover protocol.

[0062] During the soft mounting phase, the old executor retains control of the task, while the new successor silently establishes a perceptual connection with the task environment in the background and asynchronously receives the local environmental context (such as the target's trajectory history and completed progress nodes) packaged and sent by the old executor. After initializing the context state locally, the successor sends an asynchronous ready handshake signal to the old executor. Upon receiving this signal, the old executor immediately releases control and becomes idle, while the successor performs a hard switch locally, instantly taking over all physical and logical control of the role. This mechanism ensures a seamless spatial and logical transition between roles.

[0063] Step S4.3: When an agent becomes a communication island due to a communication link interruption, it performs an independent survival task according to the preset degradation rules, and automatically merges the state records through the rumor protocol when the communication network is restored, and rejoins the cooperative network.

[0064] Faced with strong electromagnetic interference or localized network disruptions caused by physical barriers, some agents may lose communication with other agents, becoming communication islands. In this embodiment, when an agent does not receive a communication message from another agent within a preset timeout period, it is determined that the agent is in a communication island state. At this time, the agent automatically suspends the asynchronous negotiation it is waiting for and obtains the remaining resource status and environmental threat status obtained locally.

[0065] The agent pre-stores degradation rules for communication island states. These rules are used to establish a correspondence between remaining resource states, environmental threat states, and independent survival tasks. In one specific embodiment, when the environmental threat state reaches a preset threat threshold (pre-set), the agent performs an in-situ concealment task; when the environmental threat state does not reach the preset threat threshold and the remaining resource state is lower than the preset resource threshold, the agent performs a silent return-to-home task; when the environmental threat state does not reach the preset threat threshold and the remaining resource state is not lower than the preset resource threshold, the agent performs a single-machine local patrol task.

[0066] After being downgraded, the agent no longer performs role tasks requiring collaboration among multiple agents; instead, it independently executes defined survival tasks. During the communication islanding period, the agent continues to record its task execution state and state changes in a locally conflict-free replicated data type structure. When the communication link is restored, the agent reconnects to the rumor protocol, merges the state increments recorded during the islanding period with the state data maintained by other agents, and rejoins the cooperative network.

[0067] This embodiment can overcome the limitations of global clock synchronization, significantly improve the communication robustness and scalability of the system, achieve low-overhead, forward-looking and low-latency response in a decentralized environment, optimize highly adaptive role scheduling and distributed collaboration capabilities in dynamic environments, and provide strong theoretical support and technical guarantee for the practical deployment of large-scale heterogeneous multi-agent systems in harsh communication networks with no center and high latency.

[0068] Furthermore, this embodiment also provides a multi-agent role autonomous negotiation and scheduling system based on asynchronous communication, including: Module M1: Asynchronous State Synchronization Module; used to build an asynchronous state synchronization model based on conflict-free copy data type and rumor protocol, and to achieve lock-free consistent updates of local states of multiple agents in a decentralized environment.

[0069] Module M2: Predictive event-driven module; used for each agent to locally maintain a perception state flow based on a sliding time window, and deploy a lightweight time-series prediction model to make real-time predictions. When the prediction result exceeds the dynamic adaptive threshold, a role handover event or a collaboration request event is generated in advance to trigger the asynchronous negotiation and scheduling process.

[0070] Module M3: Asynchronous streaming negotiation module; used to establish a streaming negotiation mechanism based on the asynchronous continuous contract network protocol, combining time decay utility function and rolling clearing strategy to complete the asynchronous dynamic allocation of roles online.

[0071] Module M4: Conflict resolution and self-healing execution module; used for distributed arbitration based on logical vector clock or hash value and rollback of conflict-free replicated data type state; uses soft mounting and hard switching to achieve seamless role handover; in isolated situations, it degrades to perform survival tasks, and rejoins the network after the network is restored by merging the state through the rumor protocol.

[0072] The module M1 includes the following sub-modules: Module M1.1: Uses a logical vector clock to assign partial-order timestamps to local events of the agent, eliminating the multi-agent system's dependence on the global physical clock.

[0073] Module M1.2: Models the role occupancy state and task scheduling queue of multiple agents as a conflict-free copy data type, supporting agents to update state data locally without blocking.

[0074] Module M1.3: Based on the rumor protocol, the state increment is asynchronously propagated within the local communication radius of the agent, and historical conflicts are resolved through mathematical merging operations to achieve eventual consistency of the network state.

[0075] Module M2 includes the following sub-modules: Module M2.1: The agent maintains a perception state stream locally based on a sliding time window. The state stream includes its own resource consumption rate, target movement trajectory, and environmental threat level.

[0076] Module M2.2: Deploys a lightweight time-series prediction model at the edge to predict the probability of task failure or the time of resource exhaustion within a set future time window in real time.

[0077] Module M2.3: When the predicted probability or prediction time exceeds the dynamic adaptive threshold, a role handover event or collaboration request event is generated in advance, transforming the delayed response into a forward-looking schedule and triggering the asynchronous negotiation process.

[0078] The module M3 includes the following sub-modules: Module M3.1: The task initiator broadcasts task feature information, carrying a logical vector clock as a timestamp and a negotiated deadline, to the local network in a publish / subscribe mode.

[0079] Module M3.2: The bidding agent calculates the utility value of executing the task by combining local execution capability with a time decay factor, wherein the time decay factor causes the utility value of the delayed response to decrease dynamically as physical time passes.

[0080] Module M3.3: The task initiator adopts a rolling clearing decision mechanism that does not wait for all nodes to bid. When the highest bid utility value received within the set rolling time window exceeds the satisfaction threshold, a confirmation message is sent asynchronously to achieve streaming task matching.

[0081] The module M4 includes the following sub-modules: Module M4.1: When multiple agent nodes concurrently compete for the same role and there is a state conflict, deterministic distributed arbitration is performed based on the partial order relationship of the vector clock or the size of the local hash value. The loser of the arbitration is automatically demoted and rolled back to its local conflict-free replicated data type state.

[0082] Module M4.2: When the old and new roles are handed over, a mechanism combining soft mounting and hard switching is adopted. The successor is in an asynchronous synchronous environment context in the background. After the asynchronous handshake is confirmed, the seamless handover of control is completed instantly.

[0083] Module M4.3: When the duration of the communication interruption between the agent and other agents reaches a preset timeout period, it determines that the agent is in a communication island state; based on the remaining resource status and environmental threat status obtained locally, it determines and executes an independent survival task according to a preset degradation rule; and merges the status records through the rumor protocol and rejoins the cooperative network when the communication network is restored.

[0084] Based on the above description of the embodiments, those skilled in the art will understand that the multi-agent role self-negotiation and scheduling method and system based on asynchronous communication described in this embodiment can be implemented in pure software or deployed and run on a general-purpose or dedicated computing hardware platform. Based on this essence, the technical solution of this embodiment can be specifically implemented in the form of a software product containing program instructions. This software product can be stored on various non-volatile storage media or directly deployed as a local or cloud service. The program instructions are used to cause computer devices with processing capabilities—including but not limited to personal computers, server clusters, mobile terminals, or other network devices—to execute the steps described in this embodiment.

[0085] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for autonomous negotiation and scheduling of multi-agent roles based on asynchronous communication, characterized in that, include: Construct an asynchronous state synchronization model based on conflict-free copy data type and rumor protocol to achieve lock-free consistent update of local state of multiple agents in a decentralized environment; Each agent locally maintains a perception state stream based on a sliding time window and deploys a lightweight time-series prediction model to make real-time predictions. When the prediction result exceeds the dynamic adaptive threshold, a role handover event or a collaboration request event is generated in advance to trigger the asynchronous negotiation and scheduling process. Asynchronous negotiation: Establish a streaming negotiation mechanism based on the asynchronous continuous contract network protocol, and combine time decay utility function and rolling clearing strategy to complete the asynchronous dynamic allocation of roles online; Scheduling process: Distributed arbitration is performed based on logical vector clock or hash value, and the state of conflict-free replicated data types is rolled back; soft mounting and hard switching are used to achieve seamless role handover; when isolated, survival tasks are downgraded, and the state is merged and re-entering the network after the network is restored via rumor protocol.

2. The method according to claim 1, characterized in that, The construction of the asynchronous state synchronization model based on conflict-free copy data type and rumor protocol includes: Logical vector clocks are used to assign partial-order timestamps to local events of the agent; The role occupancy state of multiple agents and the task scheduling queue are modeled as a conflict-free copy data type, which supports agents to update state data locally without blocking. Based on the rumor protocol, state increments are asynchronously propagated within the local communication radius of the agent, and historical conflicts are resolved through mathematical merging operations to achieve eventual consistency of the network state.

3. The method according to claim 2, characterized in that, The update rule for the logical vector clock is as follows: When the intelligent agent When a local event occurs, increment the local counter. When the intelligent agent When sending a message, the current logical vector clock is attached to the message message; When the intelligent agent Received from the intelligent agent When a message is received, the local clock is updated by taking the maximum value of each element, and the local counter is incremented.

4. The method according to claim 1, characterized in that, The perception state stream includes its own resource consumption rate, target movement trajectory, and environmental threat level.

5. The method according to claim 1, characterized in that, Using the perceived state flow as input, the lightweight time-series prediction model extracts time-series evolution features to predict in real time the probability of task failure or the time of resource exhaustion within a set future time window.

6. The method according to claim 1, characterized in that, In the asynchronous negotiation, a streaming negotiation mechanism based on the asynchronous continuous contract network protocol is established, including: The task initiator broadcasts task characteristic information, including a logical vector clock and a negotiated deadline, to the local network using a publish / subscribe model. The bidding agent calculates the utility value of the task by combining local execution capabilities with a time decay factor; The task initiator adopts a rolling clearing decision mechanism that does not wait for all nodes to bid. When the highest efficiency value received within the set rolling time window exceeds the satisfaction threshold, a confirmation message is sent asynchronously to achieve streaming task matching.

7. The method according to claim 1, characterized in that, The distributed arbitration based on logical vector clocks or hash values ​​includes: When multiple intelligent agent nodes concurrently compete for the same role, the causal order is determined based on the partial order relationship between two logical vector clocks; If there is a sequential relationship, the operation that occurs later is considered valid. If there is no order, the hash values ​​of the agent identifiers are sorted in descending order, and the hash value at the top wins. The loser in the arbitration automatically reverts to the state of a conflict-free copy data type.

8. The method according to claim 1, characterized in that, The scheduling process employs soft mounting and hard switching to achieve seamless role handover, including: The original executor retains control of the mission; The successor silently establishes a perceptual connection with the task environment in the background and asynchronously receives the local environment context packaged and sent by the old role executor. Control is transferred instantly after the asynchronous handshake is confirmed.

9. The method according to claim 1, characterized in that, The island-based survival mission is downgraded and includes: When the duration of the communication interruption between an agent and other agents reaches a preset timeout period, the agent is determined to be in a communication island state. The intelligent agent determines and executes independent survival tasks according to preset degradation rules based on the remaining resource status and environmental threat status obtained locally. The independent survival tasks include silent return to base, hiding in place, or single-machine local patrol.

10. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the method according to any one of claims 1-9.

Images