A multi-agent scenario self-adaptive deduction method for uncertain environment

Abstract

The application provides a multi-agent scenario self-adaptive deduction method for uncertain environment, and relates to the technical field of artificial intelligence and complex system deduction. The method comprises the following steps: based on the local observation data of a plurality of agents in an uncertain environment, the uncertainty level of each agent for global state cognition is evaluated, and a Bayesian belief network or reinforcement learning is adaptively selected according to the uncertainty level for state estimation and strategy generation, so that the global state cognition result of the uncertain environment is obtained; the global state cognition result is subjected to uncertainty quantitative analysis and risk assessment, so that the confidence interval of the global state variable and the risk index of strategy execution are obtained; meta-strategy is cooperated and coordinated based on the global state cognition result and the risk index, and the deduction result is obtained based on the obtained optimal cooperation scheme.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and complex system inference technology, and in particular to a multi-agent scenario adaptive inference method for uncertain environments. Background Technology

[0002] The operating environments of various complex systems, such as disaster early warning, financial market analysis, and unmanned swarm control, are becoming increasingly volatile, with uncertainty and dynamism becoming more prominent, posing significant challenges to traditional system simulation techniques. However, existing distributed collaborative solutions typically employ fixed communication protocols or predefined collaboration rules, lacking the ability to identify and quantitatively assess uncertainties in this dynamic evolution in real time. When the environment undergoes sudden changes, the system cannot adjust its collaboration strategy in a timely manner to adapt to the new state, easily leading to information asymmetry and goal conflicts among agents during the simulation process. This further exacerbates the contradiction between local optima and global optima: the rational decisions made by each agent under limited information and fixed rules, when superimposed, cause the overall simulation to deviate from the actual system evolution direction, ultimately significantly reducing the credibility and engineering practicality of the simulation results. Summary of the Invention

[0003] Therefore, it is necessary to provide a multi-agent scenario adaptive inference method for uncertain environments to address the above-mentioned technical problems.

[0004] This specification provides a multi-agent scenario adaptive inference method for uncertain environments, including: Acquire local observation data of multiple agents in an uncertain environment; The uncertainty level of each agent's perception of the global state is evaluated based on local observation data, and Bayesian belief network or reinforcement learning is adaptively selected according to the uncertainty level for state estimation and policy generation to obtain the global state perception results of the uncertain environment. Uncertainty quantification and risk assessment are performed on the global state cognition results to obtain the confidence intervals of global state variables and risk indicators for strategy execution; Based on the global state cognition results and risk indicators, meta-policy coordination is carried out. The global goal of the collaborative inference task is hierarchically decomposed into sub-goals with priority and constraints. Each agent proposes a sub-goal execution plan based on its own cognitive state. The execution plan is scored by combining the sub-goal completion degree, resource utilization rate and uncertainty risk of the global evaluation function. The conflicting plans identified by the scoring results are resolved based on game optimization to determine the optimal collaborative plan. The results obtained include global state perception, optimal collaborative solutions, confidence intervals, and risk indicators.

[0005] Furthermore, the adaptive selection of Bayesian belief networks or reinforcement learning for state estimation and policy generation based on the level of uncertainty includes: When the uncertainty level is lower than the preset uncertainty threshold, state estimation and policy generation are performed by Bayesian belief network as the main force and reinforcement learning as the auxiliary force; otherwise, state estimation and policy generation are performed by reinforcement learning as the main force and Bayesian belief network as the auxiliary force. In both modes, the auxiliary reasoning function of the non-dominant module is retained, and cognitive state sharing and auxiliary information interaction between agents are realized through global data bus. The preset uncertainty threshold is determined based on the sample mean and standard deviation of the cognitive entropy distribution obtained from historical data statistics. The sum of the product of the sample mean, the adaptive adjustment coefficient, and the standard deviation is set as the initial threshold. During the inference process, the initial threshold is dynamically adjusted according to the frequency of environmental changes and the complexity of the task through a weighted moving average method to obtain and update the preset uncertainty threshold, so as to realize the adaptive selection of reasoning strategies based on the preset uncertainty threshold.

[0006] Furthermore, the state estimation and policy generation based on Bayesian belief networks supplemented by reinforcement learning includes: A dynamic Bayesian network is constructed with local observation features and historical cognitive states as input nodes and global state probability distribution as output nodes, and the network topology is initialized by combining expert experience and data-driven approaches. The local observation features and historical cognitive states are input into the dynamic Bayesian network, and the posterior probability distribution of the global state is obtained through inference by a sampling algorithm. When new observation data is input, the conditional probability table in the dynamic Bayesian network is dynamically updated based on the new observation data to achieve real-time correction of the global state perception and obtain the global state perception result.

[0007] Furthermore, the state estimation and policy generation based on reinforcement learning supplemented by Bayesian belief networks includes: Based on the definition of an uncertain environment, a reward function containing an uncertainty penalty term is defined. The reward function is the weighted sum of the global objective benefit and the local task benefit, minus the penalty value corresponding to the uncertainty penalty term. Based on the aforementioned reward function, a fusion architecture combining deep Q-networks and proximal policy optimization algorithms is employed for policy generation and optimization. During the policy generation process, features are extracted through a shared feature extraction layer, and action value estimates and policy probability distributions are output through independent branches. The two are then weighted and fused to obtain the final action policy. During the strategy optimization process, a priority experience replay mechanism is introduced. This mechanism assigns differentiated sampling priorities to historical experience samples based on the magnitude of the temporal difference error in the strategy evaluation. This prioritizes the sampling and learning of experience in high-uncertainty scenarios, thereby completing the strategy optimization and obtaining the global state cognition result.

[0008] Furthermore, the assessment of the uncertainty level of each agent's perception of the global state based on local observation data includes: Infer the probability distribution of each global state hypothesis based on local observation data; Information entropy is obtained based on the probability distribution of the global state assumption. The more dispersed the probability distribution, the greater the information entropy and the higher the level of uncertainty in the representation.

[0009] Furthermore, the uncertainty quantification analysis and risk assessment of the global state cognition results, to obtain the confidence intervals of the global state variables and risk indicators for strategy execution, include: The global state perception results are used to extract uncertainty features through feature engineering to distinguish different sources of information loss, noise interference, and rule ambiguity. Based on the uncertainty characteristics, a quantization model is constructed using Monte Carlo sampling or Bayesian deep learning. The global state variable is then randomly sampled multiple times using the quantization model to generate a dataset. Statistical features are calculated based on the dataset to generate confidence intervals. Based on the dataset, the failure probability and expected loss of the strategy execution are calculated using a pre-set risk assessment model to obtain risk indicators.

[0010] Furthermore, the meta-strategy coordination based on global state perception results and risk indicators also includes: After optimizing the conflict resolution scheme and determining the optimal collaborative scheme through game theory, the core decision-making authority is dynamically allocated through a weighted voting algorithm based on the cognitive accuracy, resource status and task completion of each agent, so that the key sub-objectives are led by the agent with the highest reliability. Furthermore, based on the global inference results, the collaborative strategy parameters are updated periodically to optimize the interaction rules between agents and achieve meta-policy collaborative coordination.

[0011] Furthermore, before acquiring local observation data of multiple agents in an uncertain environment, a parameter initialization operation is also included, including: Configure the number of agents, initial cognitive state, global inference target and uncertainty quantification threshold. The uncertainty quantification threshold is calculated based on Monte Carlo sampling to measure the uncertainty of global state cognition and is determined by setting a confidence interval width threshold. After obtaining the deduction results, which include global state cognition results, optimal collaborative schemes, confidence intervals, and risk indicators, the process also includes an iterative judgment step: Determine whether the termination conditions are met. The termination conditions include the achievement of the global goal, the uncertainty quantification index being lower than a preset threshold, or the maximum number of simulation iterations being reached. If the termination conditions are not met, return to the step of obtaining local observation data and continue iterative execution.

[0012] Furthermore, acquiring local observation data of multiple agents in an uncertain environment includes: Sensor data, historical experience data, and third-party heterogeneous data are collected through multi-source data fusion and then packaged in a standardized manner. An adaptive algorithm is used to perform hierarchical filtering of noisy data on the standardized and packaged data, and a time-series interpolation model based on an attention mechanism is used to complete the missing data. Based on the completed data, a lightweight convolutional neural network or a temporal attention network is used to extract key features of local observations, and the obtained high-dimensional feature vectors are used as local observation data.

[0013] Furthermore, the process of resolving conflicting solutions identified based on the scoring results using game theory optimization to determine the optimal collaborative solution includes: A two-agent zero-sum game model is constructed and a payoff matrix is ​​defined. The payoff matrix is ​​calculated by weighting and comprehensively calculating the payoff of task completion, resource consumption cost and risk cost to obtain the agent utility value under each strategy combination. Based on the payoff matrix, the Lemke-Howson algorithm is used to iteratively solve the Nash equilibrium strategy. The Lemke-Howson algorithm starts from the initial feasible strategy profile and iteratively adjusts the mixed strategy until the equilibrium state is reached. As the number of agents increases, a hierarchical approximation method is used to decompose the multi-agent game into a sequence of two-agent subgames to determine the optimal cooperative solution.

[0014] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: This invention adaptively selects either a Bayesian belief network or reinforcement learning for state estimation and policy generation based on the uncertainty level of the agent's perception of the global state. This allows the system to efficiently reason using a Bayesian network when the environment is relatively stable, and switches to reinforcement learning for adaptive policy generation when the environment changes abruptly and uncertainty increases. This improves the cognitive robustness and adaptability to unsteady states in uncertain environments. Simultaneously, the uncertainty of the global state perception is quantified, providing a reliable confidence boundary for collaborative decision-making. Information asymmetry leads to each agent constructing its own beliefs based only on local observations, resulting in a deviation between its individual objective function and the system's global objective. By explicitly introducing an uncertainty risk constraint factor, the system addresses this deviation. The uncertainty of an agent's perception of the global state is quantified into an optimizable penalty term and incorporated into the global evaluation function. This ensures that any individual behavior that deviates from the global objective is effectively suppressed due to the introduction of additional risk costs, thereby breaking the non-cooperative game structure under information asymmetry. By solving the joint strategy that maximizes the global evaluation function, each agent is forced to move from its local optimum towards the global optimization objective of the system. Ultimately, under the dual effects of risk sharing and benefit coordination, strict alignment between individual behavior and the global objective is achieved. The final output includes a deduction result containing multi-dimensional quantitative indicators and the optimal collaborative solution, improving the reliability, collaborative consistency, and practicality of engineering decision-making for complex systems under highly uncertain environments. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0016] Figure 1 This document provides a flowchart illustrating a multi-agent scenario adaptive inference method for uncertain environments. Figure 2 This document provides a schematic diagram of the overall architecture of a multi-agent scenario adaptive inference system. Figure 3 This is a schematic diagram of an adaptive switching mechanism for dual cognitive modules provided in this specification. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0018] The technical solution provided by this invention can be applied to various complex system simulation scenarios, such as disaster early warning simulation, financial market risk analysis, unmanned swarm task planning, and emergency management decision support. Addressing the technical shortcomings of existing simulation systems, such as drastic performance degradation under highly uncertain environments, lack of effective coordination mechanisms among multiple agents, and insufficient uncertainty quantification, this invention provides a multi-agent scenario-adaptive simulation engine and its implementation method for uncertain environments, improving the reliability and adaptability of simulations for complex systems in unsteady environments.

[0019] The following is combined with Figure 1 This invention describes a multi-agent scenario adaptive inference method for uncertain environments.

[0020] In practical implementation, this method can be implemented based on an overall architecture of "layered decoupling + distributed collaboration". Figure 2 This is a schematic diagram of the overall architecture of the multi-agent scenario adaptive inference system provided in this specification, as shown below. Figure 2 As shown, the system is divided into a perception layer, a cognitive decision-making layer, a coordination layer, and an inference output layer from top to bottom. A global data bus and an uncertainty management center are also introduced to achieve efficient linkage between modules and full-process control of uncertain information. The overall architecture is modular and scalable, supporting intelligent agent configuration adaptation and functional customization for different application scenarios. The implementation of this invention will be described in detail below with specific steps.

[0021] S1. Acquire local observation data of multiple agents in an uncertain environment.

[0022] For example, S1 is completed through the perception layer. The perception layer is responsible for environmental data acquisition, preprocessing, and the generation of local observation information, addressing information gaps and noise interference, and providing reliable data input for subsequent cognitive decision-making. Specifically, it includes the following sub-steps:

[0023] First, a multi-source data fusion acquisition unit acquires sensor data (such as disaster monitoring sensors, financial data interfaces, and unmanned system sensors), historical experience data, and heterogeneous third-party data, and encapsulates them uniformly using a standardized data format (such as JSON). Second, a noise filtering and missing value completion unit uses an adaptive algorithm to perform hierarchical filtering of noisy data and employs an attention-based temporal interpolation model to complete missing data. Then, a local observation feature extraction unit uses a lightweight convolutional neural network (CNN) or a temporal attention network to extract key features of local observations, generating high-dimensional feature vectors as local observation data for subsequent steps.

[0024] It should be noted that before acquiring local observation data, parameter initialization operations are also included: configuring the number of multi-agents, the initial cognitive state, the global inference objective, and the uncertainty quantification threshold. The uncertainty quantification threshold is determined by calculating the uncertainty measure of the global state cognition based on Monte Carlo sampling and by setting a confidence interval width threshold.

[0025] S2. Evaluate the uncertainty level of each agent's perception of the global state based on local observation data, and adaptively select Bayesian belief network or reinforcement learning to perform state estimation and policy generation according to the uncertainty level, so as to obtain the global state perception result of the uncertain environment.

[0026] For example, S2 is executed by the cognitive decision-making layer. Each agent is configured with switchable cognitive submodules: a Bayesian belief network submodule and a reinforcement learning submodule, to achieve a closed loop of cognition, correction, and decision-making.

[0027] Figure 3 This is a schematic diagram of an adaptive switching mechanism for a dual-cognitive module provided in this specification, such as... Figure 3 As shown, the uncertainty level is first assessed. Based on local observation data, the probability distribution of each global state hypothesis is inferred. Information entropy, or cognitive entropy, is obtained from the probability distribution. The more dispersed the probability distribution, the greater the information entropy, and the higher the level of uncertainty it represents. The formula for calculating cognitive entropy is... in, Represents cognitive entropy. Indicates the first The probability values ​​of each global state hypothesis. This represents the total number of state assumptions.

[0028] Secondly, the cognitive sub-module is adaptively selected based on the level of uncertainty. The preset uncertainty threshold is determined based on the sample mean and standard deviation of the cognitive entropy distribution obtained from historical data statistics: during the initialization phase, samples of the cognitive entropy distribution under different environmental conditions are collected, and the sample mean is calculated. and standard deviation An adaptive threshold setting formula is adopted. Set initial threshold ( The value ranges from 1.5 to 3.0, with an initial threshold that can be set to 0.7. During the simulation, the threshold is dynamically adjusted using a weighted moving average method based on the frequency of environmental changes and the complexity of the task. The expression is as follows: ,in The adaptive learning rate (range 0.1~0.3).

[0029] The specific rules for adaptive selection are as follows: when the cognitive entropy is below a preset threshold, i.e., the environment is relatively stable and the information is relatively complete, the Bayesian belief network is used as the primary driver, supplemented by reinforcement learning, for state estimation and policy generation to ensure inference efficiency; when the cognitive entropy is above the preset threshold, i.e., the environment is highly uncertain and information is severely lacking, the system switches to reinforcement learning as the primary driver, supplemented by the Bayesian belief network, for state estimation and policy generation to improve the adaptive capability of the policy. In both modes, the auxiliary reasoning function of the non-dominant module is retained, and cognitive state sharing and auxiliary information interaction among agents are achieved through a global data bus.

[0030] Specifically, the implementation method when Bayesian belief networks are dominant is as follows: Construct a dynamic Bayesian network, using local observation features and historical cognitive states as input nodes and the global state probability distribution as output nodes. Initialize the network topology using a combination of expert experience and data-driven approaches. Input the local observation features and historical cognitive states into the dynamic Bayesian network, and use a sampling algorithm for inference to obtain the posterior probability distribution of the global state. When new observation data is input, dynamically update the conditional probability table based on the new observation data to achieve real-time correction of the global state cognition and obtain the global state cognition result.

[0031] The implementation method when reinforcement learning is the primary approach is as follows: First, define the state space (local observation features + global cognitive state fragments) and the action space (a set of inference strategies, such as evacuation route adjustments in disaster early warning or task allocation schemes in unmanned swarms). Based on the uncertain environment, define a reward function that includes an uncertainty penalty term, with the expression: ,in As a global objective reward, For local task rewards, For uncertainty penalty items, , These are the weighting coefficients.

[0032] Then, a fusion architecture combining Deep Q-Network (DQN) and Proximal Policy Optimization (PPO) is used for policy generation and optimization. During policy generation, features are extracted through a shared feature extraction layer (which can be a 3-layer fully connected neural network with 256, 128, and 64 neurons, using ReLU activation function). These features are then extracted through independent branches: action value estimation (DQN branch output layer has the same number of neurons as the action space dimension, with a linear activation function) and policy probability distribution (PPO branch output layer uses softmax activation function). Finally, the action policy is obtained through weighted fusion. Among them, the fusion weight =0.4, =0.6, which can be adaptively adjusted according to the training phase.

[0033] During policy optimization, a Priority Experience Replay (PER) mechanism is introduced. PER assigns differentiated sampling priorities to historical experience samples based on the magnitude of the temporal difference error in policy evaluation. The temporal difference error is calculated using the following formula: The priority calculation function is: , Let's take 0.01. The sampling probability formula is: It prioritizes sampling and learning from experience in high-uncertainty scenarios to optimize policies and obtain global state cognition results. It also supports online incremental updates of model parameters and triggers an emergency learning mechanism to quickly adjust policies when environmental changes occur.

[0034] S3. Conduct uncertainty quantification analysis and risk assessment on the global state cognition results to obtain the confidence interval of the global state variables and the risk indicators of strategy execution.

[0035] For example, S3 is executed by the Uncertainty Management Center to achieve unified control over the entire uncertainty process. This includes identifying sources of uncertainty, building quantitative models, and generating indicators.

[0036] First, an uncertainty source identification unit extracts uncertainty features using feature engineering, distinguishing different sources such as missing information, noise interference, and rule ambiguity. Then, based on these uncertainty features, a quantification model is constructed using Monte Carlo sampling or Bayesian deep learning. This model is then used to generate a dataset by repeatedly sampling global state variables randomly. For example, Monte Carlo simulation can be used to sample uncertainties in the deduction process multiple times (e.g., 1000 times), tracking the propagation path and impact of uncertainty in the deduction chain and generating an uncertainty propagation map. Finally, statistical features (mean, variance, etc.) are calculated based on the dataset, generating confidence intervals (confidence level is configurable, default 95%). Simultaneously, a pre-defined risk assessment model is used to calculate the failure probability and expected loss of the strategy execution, obtaining risk indicators.

[0037] S4. Based on the global state cognition results and risk indicators, meta-policy coordination is carried out. The global goal of the collaborative simulation task is hierarchically decomposed into sub-goals with priority and constraints. Each agent proposes a sub-goal execution plan based on its own cognitive state. The execution plan is scored by combining the sub-goal completion degree, resource utilization rate and uncertainty risk of the global evaluation function. The conflicting plans identified by the scoring results are resolved by game optimization to determine the optimal collaborative plan.

[0038] For example, S4 is executed by the coordination layer. The coordination layer uses a meta-policy coordination mechanism to achieve goal alignment, behavioral coordination, and conflict resolution among multiple agents, ensuring consistency in global inference.

[0039] First, based on the needs of fields such as "maximizing personnel evacuation efficiency" in disaster early warning and "maximizing task completion rate" in unmanned swarms, a hierarchical task planning algorithm decomposes the global objective into allocable sub-objectives, generating sub-objective priorities and constraints, such as resource limitations and time windows. Then, a distributed collaborative negotiation mechanism of "local proposal-global evaluation-conflict resolution" is adopted. Each agent proposes a sub-objective execution plan based on its own cognitive state, and the plan is scored by a global evaluation function that considers sub-objective completion, resource utilization, and uncertainty risk. For conflicting plans identified by the scoring results, such as resource competition and path overlap, game theory optimization is used to resolve them to determine the optimal collaborative solution.

[0040] The specific implementation of game optimization includes: constructing a two-agent zero-sum game model and defining a payoff matrix. The payoff matrix weights and comprehensively calculates the payoff based on task completion, resource consumption, and risk, yielding the agent's utility value for each strategy combination. For example, payoff matrix elements... ,in Indicates the reward for task completion. This indicates a penalty for resource consumption. This represents the risk penalty term. Based on the payoff matrix, the Lemke-Howson algorithm is used iteratively to solve for the Nash equilibrium strategy. The algorithm starts from an initial feasible strategy profile and iteratively adjusts the mixed strategy until an equilibrium state is reached (the maximum number of iterations is 1000 by default). When the number of agents increases, a hierarchical approximation method is used to decompose the multi-agent game into a sequence of two-agent subgames, reducing the computational complexity to... ,in The number of subgames is used to determine the optimal cooperative solution.

[0041] After optimizing conflict resolution strategies and determining the optimal collaborative approach through game theory, the process also includes dynamic role and permission allocation: based on each agent's cognitive accuracy, resource status, and task completion performance, core decision-making permissions are dynamically allocated using a weighted voting algorithm, ensuring that key sub-objectives are led by the agent with the highest reliability. Simultaneously, collaborative strategy parameters are periodically updated based on global projection results (such as task completion rate and risk control effectiveness) to optimize interaction rules between agents and achieve meta-policy collaborative coordination.

[0042] S5. Obtain the deduction results, which include global state cognition results, optimal collaborative solutions, confidence intervals, and risk indicators.

[0043] For example, S5 is completed by the inference output layer. The inference output layer is responsible for outputting the integrated inference-related information and providing decision-makers with a robust basis for judgment.

[0044] The specific output includes: the corrected global state perception results, the optimal collaborative solution (i.e., the optimized inference strategy), confidence intervals, and risk assessment indicators. A visual interface can be provided to display the inference process, global state changes, agent behavior trajectories, and uncertainty quantification results (such as confidence interval curves and risk heatmaps). Decision-makers can adjust the inference parameters and view the adjusted inference results.

[0045] After obtaining the simulation results, an iterative determination step is included: determining whether the termination conditions are met. Termination conditions include: global goal achievement, uncertainty quantification index falling below a preset threshold (i.e., cognitive reliability meeting the quantification threshold, such as confidence interval width being less than the set threshold), or reaching the maximum number of simulation iterations. If the termination conditions are not met, the process returns to step S1 (obtaining local observation data) to continue iterative execution; if they are met, the simulation ends.

[0046] To more clearly illustrate the overall implementation process of this invention, the following example uses a flood disaster early warning simulation in a certain region. Assume that three agents need to be coordinated simultaneously: Agent A is responsible for the upstream monitoring station area, Agent B is responsible for the midstream residential area, and Agent C is responsible for the downstream flood control facilities. The overall objective is to "achieve zero casualties and minimize property damage within 2 hours."

[0047] Parameter initialization phase (S1 pre-operation): Configure 3 agents, with the initial cognitive state being that the risk level of each area is "low". The global inference objective is the above-mentioned multi-objective optimization problem, and the uncertainty quantification threshold is set to a 95% confidence interval width of less than 0.15.

[0048] Local Observation Acquisition Phase (S1): Each agent collects local data through the perception layer—Agent A acquires upstream water level sensor data: current water level 52.3m, trending upward 0.2m / h, meteorological radar data, and historical flood peak records; Agent B acquires residential population density data: 1200 people / km², evacuation route status, and shelter capacity; Agent C acquires downstream dam safety factor: 0.85, flood control facility operation status, and emergency material reserves. After multi-source fusion, noise filtering, and feature extraction, each agent generates a local observation feature vector, such as 64-dimensional sensor data + 16-dimensional temporal features.

[0049] Cognitive Iteration and Correction Phase (S2): In the first round of inference, the cognitive entropy was calculated to be 0.92, lower than the preset initial threshold of 0.7, therefore the Bayesian Belief Network submodule took the lead. Reasoning was performed through a dynamic Bayesian network, outputting the risk probability distribution for each region: upstream flood peak probability 35%, midstream waterlogging probability 28%, and downstream dike breach probability 12%. In the second round of inference, an emergency warning was received from the meteorological bureau stating "heavy rainstorms are expected in the basin within the next 3 hours." After updating the observation data, the cognitive entropy rose to 1.35, exceeding the threshold, and the reinforcement learning submodule was automatically switched to take the lead. A DQN+PPO fusion algorithm was adopted. The state space was an 80-dimensional vector containing features such as water level, population distribution, and evacuation progress in each region. The action space contained eight actions, such as "upstream flood discharge," "midstream evacuation," and "downstream dike reinforcement," and their combinations. The reward function was set with an uncertainty penalty coefficient β=0.5. Finally, the optimal action strategy was obtained through weighted fusion.

[0050] Uncertainty Quantification Analysis Phase (S3): Through uncertainty source identification, the model building unit outputs the following uncertainties: upstream water level prediction uncertainty contributes 42%, weather forecast uncertainty contributes 35%, and agent cognitive bias contributes 23%. Monte Carlo sampling (1000 times) is used to calculate confidence intervals: the 95% confidence interval for upstream water level is [52.3m, 56.8m], and the 95% confidence interval for midstream evacuation time is [45min, 78min]. The risk assessment model calculates the failure probability of the strategy execution: the failure probability of the "full evacuation strategy" is 8%, and the failure probability of the "partial evacuation strategy" is 23%.

[0051] Meta-strategy Coordination Phase (S4): The global objective is decomposed into three sub-objectives—upstream sub-objective 1: maintain dam safety, priority P1=0.9; midstream sub-objective 2: complete personnel evacuation within 2 hours, priority P2=1.0; downstream sub-objective 3: minimize property loss, priority P3=0.7. After each agent proposes a solution, resource allocation conflicts are identified through a global evaluation function. For example, personnel evacuation requires competition between opening shelters and dam reinforcement resources. The Lemke-Howson algorithm is used to solve the 2×2 payoff matrix, yielding the equilibrium strategy: prioritize ensuring midstream evacuation resources, and adopt a conservative approach for downstream dam reinforcement. Dynamic role allocation is adjusted based on real-time cognitive accuracy, temporarily assigning core decision-making authority to agent B with the highest cognitive accuracy.

[0052] Output and Iterative Judgment Phase (S5 and Subsequent Judgments): The output deduction result is the optimal strategy—"Immediately initiate the evacuation of midstream residential areas, open the upstream floodgates to release 20% of the floodwater to alleviate pressure, and maintain the current protection level downstream"; under the 95% confidence interval, the expected casualties are 0-2 people, and the expected property losses are 5-12 million yuan. The iterative judgment finds that the cognitive entropy has decreased to 0.68 (below the threshold), the uncertainty quantification index meets the requirements, the termination condition is met, and the deduction ends.

[0053] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0054] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A multi-agent scenario adaptive inference method for uncertain environments, characterized in that, include: Acquire local observation data of multiple agents in an uncertain environment; The uncertainty level of each agent's perception of the global state is evaluated based on local observation data, and Bayesian belief network or reinforcement learning is adaptively selected according to the uncertainty level for state estimation and policy generation to obtain the global state perception results of the uncertain environment. Uncertainty quantification and risk assessment are performed on the global state cognition results to obtain the confidence intervals of global state variables and risk indicators for strategy execution; Based on the global state cognition results and risk indicators, meta-policy coordination is carried out. The global goal of the collaborative inference task is hierarchically decomposed into sub-goals with priority and constraints. Each agent proposes a sub-goal execution plan based on its own cognitive state. The execution plan is scored by combining the sub-goal completion degree, resource utilization rate and uncertainty risk of the global evaluation function. The conflicting plans identified by the scoring results are resolved based on game optimization to determine the optimal collaborative plan. The results obtained include global state perception, optimal collaborative solutions, confidence intervals, and risk indicators.

2. The multi-agent scenario adaptive inference method for uncertain environments as described in claim 1, characterized in that, The adaptive selection of Bayesian belief networks or reinforcement learning for state estimation and policy generation based on the level of uncertainty includes: When the uncertainty level is lower than the preset uncertainty threshold, state estimation and policy generation are performed by Bayesian belief network as the main force and reinforcement learning as the auxiliary force; otherwise, state estimation and policy generation are performed by reinforcement learning as the main force and Bayesian belief network as the auxiliary force. In both modes, the auxiliary reasoning function of the non-dominant module is retained, and cognitive state sharing and auxiliary information interaction between agents are realized through global data bus. The preset uncertainty threshold is determined based on the sample mean and standard deviation of the cognitive entropy distribution obtained from historical data statistics. The sum of the product of the sample mean, the adaptive adjustment coefficient, and the standard deviation is set as the initial threshold. During the inference process, the initial threshold is dynamically adjusted according to the frequency of environmental changes and the complexity of the task through a weighted moving average method to obtain and update the preset uncertainty threshold, so as to realize the adaptive selection of reasoning strategies based on the preset uncertainty threshold.

3. The multi-agent scenario adaptive inference method for uncertain environments as described in claim 2, characterized in that, The method of using Bayesian belief networks as the primary approach, supplemented by reinforcement learning, for state estimation and policy generation includes: A dynamic Bayesian network is constructed with local observation features and historical cognitive states as input nodes and global state probability distribution as output nodes, and the network topology is initialized by combining expert experience and data-driven approaches. The local observation features and historical cognitive states are input into the dynamic Bayesian network, and the posterior probability distribution of the global state is obtained through inference by a sampling algorithm. When new observation data is input, the conditional probability table in the dynamic Bayesian network is dynamically updated based on the new observation data to achieve real-time correction of the global state perception and obtain the global state perception result.

4. The multi-agent scenario adaptive inference method for uncertain environments as described in claim 2, characterized in that, The method of using reinforcement learning as the primary method, supplemented by Bayesian belief networks for state estimation and policy generation, includes: Based on the definition of an uncertain environment, a reward function containing an uncertainty penalty term is defined. The reward function is the weighted sum of the global objective benefit and the local task benefit, minus the penalty value corresponding to the uncertainty penalty term. Based on the aforementioned reward function, a fusion architecture combining deep Q-networks and proximal policy optimization algorithms is employed for policy generation and optimization. During the policy generation process, features are extracted through a shared feature extraction layer, and action value estimates and policy probability distributions are output through independent branches. The two are then weighted and fused to obtain the final action policy. During the strategy optimization process, a priority experience replay mechanism is introduced. This mechanism assigns differentiated sampling priorities to historical experience samples based on the magnitude of the temporal difference error in the strategy evaluation. This prioritizes the sampling and learning of experience in high-uncertainty scenarios, thereby completing the strategy optimization and obtaining the global state cognition result.

5. The multi-agent scenario adaptive inference method for uncertain environments as described in claim 1, characterized in that, The assessment of the uncertainty level of each agent's perception of the global state based on local observation data includes: Infer the probability distribution of each global state hypothesis based on local observation data; Information entropy is obtained based on the probability distribution of the global state assumption. The more dispersed the probability distribution, the greater the information entropy and the higher the level of uncertainty in the representation.

6. The multi-agent scenario adaptive inference method for uncertain environments as described in claim 1, characterized in that, The uncertainty quantification analysis and risk assessment of the global state perception results, to obtain the confidence intervals of the global state variables and risk indicators for strategy execution, include: The global state perception results are used to extract uncertainty features through feature engineering to distinguish different sources of information loss, noise interference, and rule ambiguity. Based on the uncertainty characteristics, a quantization model is constructed using Monte Carlo sampling or Bayesian deep learning. The global state variable is then randomly sampled multiple times using the quantization model to generate a dataset. Statistical features are calculated based on the dataset to generate confidence intervals. Based on the dataset, the failure probability and expected loss of the strategy execution are calculated using a pre-set risk assessment model to obtain risk indicators.

7. The multi-agent scenario adaptive inference method for uncertain environments as described in claim 1, characterized in that, The meta-strategy coordination based on global state perception results and risk indicators also includes: After optimizing the conflict resolution scheme and determining the optimal collaborative scheme through game theory, the core decision-making authority is dynamically allocated through a weighted voting algorithm based on the cognitive accuracy, resource status and task completion of each agent, so that the key sub-objectives are led by the agent with the highest reliability. Furthermore, based on the global inference results, the collaborative strategy parameters are updated periodically to optimize the interaction rules between agents and achieve meta-policy collaborative coordination.

8. The multi-agent scenario adaptive inference method for uncertain environments as described in claim 1, characterized in that, Before acquiring local observation data of multiple agents in an uncertain environment, the method further includes parameter initialization operations, including: Configure the number of agents, initial cognitive state, global inference target and uncertainty quantification threshold. The uncertainty quantification threshold is calculated based on Monte Carlo sampling to measure the uncertainty of global state cognition and is determined by setting a confidence interval width threshold. After obtaining the deduction results, which include global state cognition results, optimal collaborative schemes, confidence intervals, and risk indicators, the process also includes an iterative judgment step: Determine whether the termination conditions are met. The termination conditions include the achievement of the global goal, the uncertainty quantification index being lower than a preset threshold, or the maximum number of simulation iterations being reached. If the termination conditions are not met, return to the step of obtaining local observation data and continue iterative execution.

9. The multi-agent scenario adaptive inference method for uncertain environments as described in claim 7, characterized in that, The acquisition of local observation data of multiple agents in an uncertain environment includes: Sensor data, historical experience data, and third-party heterogeneous data are collected through multi-source data fusion and then packaged in a standardized manner. An adaptive algorithm is used to perform hierarchical filtering of noisy data on the standardized and packaged data, and a time-series interpolation model based on an attention mechanism is used to complete the missing data. Based on the completed data, a lightweight convolutional neural network or a temporal attention network is used to extract key features of local observations, and the obtained high-dimensional feature vectors are used as local observation data.

10. The multi-agent scenario adaptive inference method for uncertain environments as described in claim 1, characterized in that, The process of resolving conflicting solutions identified based on scoring results using game theory optimization to determine the optimal collaborative solution includes: A two-agent zero-sum game model is constructed and a payoff matrix is ​​defined. The payoff matrix is ​​calculated by weighting and comprehensively calculating the payoff of task completion, resource consumption cost and risk cost to obtain the agent utility value under each strategy combination. Based on the payoff matrix, the Lemke-Howson algorithm is used to iteratively solve the Nash equilibrium strategy. The Lemke-Howson algorithm starts from the initial feasible strategy profile and iteratively adjusts the mixed strategy until the equilibrium state is reached. As the number of agents increases, a hierarchical approximation method is used to decompose the multi-agent game into a sequence of two-agent subgames to determine the optimal cooperative solution.

Images