Platform and method for analyzing causal effect of multi-agent system

Abstract

The invention discloses a multi-agent system causal effect analysis platform and method, and the platform comprises a data processing module, an objective world model, a subject world model, a causal intervention module, an intention detection and clustering module, an active experiment planning module, and a visualization and report module. The causal intervention module comprises an observation analysis unit, an intervention analysis unit and a mechanism analysis unit; the observation and analysis unit is used for identifying a state space, an event sequence and probability distribution of a complex application field under natural evolution, and constructing a potential cause and effect graph structure; the intervention analysis unit quantifies the real causal effect output by intervention to the application field by applying exogenous intervention to the key variables; the mechanism analysis unit goes deep into a world model in a single agent and tracks a thinking path, a decision intention and an interaction rule of a microscopic agent individual; the method is suitable for a platform for analyzing a multi-agent system, and effect estimation and mechanism explanation are achieved.

Description

Technical Field

[0001] This invention relates to the field of computational experiments and causal reasoning technology for complex systems, and in particular to a platform and method for causal effect analysis of multi-agent systems. Background Technology

[0002] In the study of complex systems, the identification and interpretation of causal relationships has always been a significant challenge for the scientific community. Complex systems typically consist of a large number of interacting individuals, whose microscopic dynamics are often disturbed by noise, resulting in relatively weak causal relationships at this scale. The "causal emergence" theory proposed by Hoel et al. posits that macroscopic causal relationships are more stable and identifiable than those at the microscopic level. Microscopic individuals emerge macroscopic properties through regular interactions, and these macroscopic properties cannot be directly derived from a single element but are the result of the complex interactions of the entire system. Therefore, although many studies tend to examine causal effects at the macroscopic level, macroscopic analysis struggles to reveal the generative processes of causal mechanisms, limiting the understanding of emergent phenomena.

[0003] To address this, academia has gradually developed two paradigms of causal analysis: "horizontal causality" and "vertical causality." Horizontal causality focuses on the correlation between variables, representing a dependent causal perspective that emphasizes the identification and quantification of causal effects. Its goal is to discover whether a causal relationship exists between variables and to estimate the strength of this causal effect. This analysis typically utilizes statistical models, such as structural equation modeling, causal graphs, and potential outcome analysis, and emphasizes data quality and the robustness of statistical inference. In contrast, vertical causality emphasizes the generative path of causal mechanisms, belonging to a generative causal paradigm. It focuses on how one individual influences another, and how this influence, through interaction, triggers collective phenomena at the system level. Vertical causality advocates revealing the formation mechanism from cause to effect by dissecting the system structure and simulating its dynamic interaction process, thus placing greater emphasis on the mechanistic explanation of the process. Especially in social and economic systems, ABM (multi-agent modeling) has become an important tool for achieving vertical causal analysis. Based on individual attributes and interaction rules, it gradually constructs explanations of the paths leading to macroscopic phenomena by simulating and reproducing emergent processes.

[0004] However, both approaches have unavoidable limitations. While horizontal causality can identify causal relationships at the data level, it cannot reveal causal pathways; while vertical causality excels at revealing mechanisms, it is difficult to directly verify its real-world validity. Therefore, in the study of complex systems, the combination of horizontal and vertical causality is becoming increasingly important. Horizontal analysis provides experimental design basis and data support for vertical modeling, while vertical simulation can verify whether the causal structure revealed by horizontal analysis possesses generative properties, thus achieving a complementary fusion of causal inference and causal explanation.

[0005] Beyond the perspectives of horizontal and vertical causality, causal issues can also be categorized into "positive causality" and "negative causality" based on research paths. The former studies "what would happen if X were done," emphasizing the prediction of the consequences of interventions, and is typically used for policy evaluation and behavioral prediction. The latter, however, asks "what caused Y," focusing on the causal explanation of observed outcomes. Goldthorpe proposed three ways of understanding causality—robust dependence, outcome manipulation, and generative processes—further enriching the hierarchical understanding of causal issues. Robust dependence emphasizes causal inference under the control of confounding variables; outcome manipulation emphasizes changes in outcomes under intervention; and generative processes attempt to explain the specific paths from micro to macro phenomena. Multi-agent modeling plays a crucial role in explaining generative processes; by simulating agent behavior and interaction rules, it tracks the emergence mechanisms of macroeconomic outcomes, becoming a core tool for negative causal analysis.

[0006] In the study of causal relationships, understanding "causal effects" is not the same as understanding "causal mechanisms." Especially in complex systems, simple inferences between variables are insufficient to reconstruct the nonlinearity, temporal sequence, and structural coupling in the system's evolution. Therefore, computational experiments have emerged as a bridge connecting positive and negative causality. By simulating the evolution of an "artificial society," computational experiments can not only predict causal effects between variables but also trace micro-pathways and reconstruct the system's causal mechanisms. In this process, the world model becomes the core of computational experiments, comprising an external environmental state model and an internal agent belief model, respectively characterizing the structural laws of the real system and the agent's behavioral expectations. By constructing world models and conducting simulations, computational experiments can achieve algorithmic estimation of causal effects and reveal underlying generative mechanisms by constructing counterfactual trajectories.

[0007] Furthermore, the entire process of computational experimental causal analysis can be divided into three levels: the first level is association analysis, which answers whether a statistical association exists between variables; the second level is intervention analysis, which assesses causal effects by simulating changes in intervention variables; and the third level is mechanism analysis, which reveals how variables produce results through a series of interactive processes. This hierarchical structure not only reflects the advancement of causal research from description to explanation but also reflects the logical evolution of causal analysis in complex systems from static dependence to dynamic generation. In recent years, this framework has continuously integrated various technical approaches, such as propensity score matching, instrumental variable methods, structural equation modeling, DAG graph analysis, causal forests, Bayesian regression trees, and deep causal learning, providing methodological support for achieving systematic, interpretable, and verifiable causal inferences.

[0008] In summary, to bridge the gap between causal effect prediction and causal mechanism explanation, a comprehensive method that unifies the logic of these two types of causal analysis is urgently needed. This invention proposes a computational experimental causal analysis method for complex systems, integrating modeling approaches for both positive and negative causality. Relying on simulation experiments and world model construction techniques, it achieves dual deduction of causal structure identification and mechanism generation processes on a unified platform, providing more systematic and operational technical support for causal relationship research.

[0009] Summary of the invention.

[0010] This invention proposes a computational experimental causal analysis method for complex systems, addressing three core bottlenecks in existing causal inference techniques for complex system scenarios: first, the causal mechanism is invisible, and traditional methods can only identify variable correlations but cannot explain the generation path; second, intervention is limited, as real-world experiments are difficult to conduct due to ethical, cost, or non-reproducibility issues; and third, there is a lack of unified modeling from micro to macro levels, making it difficult to simultaneously analyze causal effects and causal mechanisms within the same framework. Based on a multi-agent artificial social model, this invention integrates key technologies such as structural causal modeling (SCM), counterfactual simulation, world model building, large language modeling (LLM), intent analysis, and dynamic intervention to form a complete multi-level causal inference system. This system enables full-chain causal analysis and verification of complex systems in a virtual environment.

[0011] The present invention adopts the following technical solution:

[0012] A platform for causal effect analysis in multi-agent systems, comprising: a data processing module, an objective world model, an agent world model, a causal intervention module, an intent detection and clustering module, an active experiment planning module, and a visualization and reporting module; the causal intervention module includes an observation analysis unit, an intervention analysis unit, and a mechanism analysis unit; wherein...

[0013] The data processing module abstracts the real world into relevant data of an artificial society or domain, and obtains panel data through computational experiments.

[0014] The objective world model is used to express the environment, structure, state transitions and variable dependencies of the application field. It is the operating vehicle of the entire experiment and is used to characterize the state transitions, variable dependencies and evolution laws driven by observation data.

[0015] The subject-world model is used to be embedded in each intelligent agent to characterize the individual's cognition, intention, learning, and decision-making logic regarding the environment;

[0016] The observation and analysis unit is used to identify the state space, event sequence, and probability distribution of complex application domains under natural evolution, and to construct a potential causal graph structure.

[0017] The intervention analysis quantifies the true causal effect of the intervention on the application field by applying exogenous intervention to key variables and conducting parallel simulations between the treatment group and the control group.

[0018] The mechanism analysis unit delves into the world model within a single intelligent agent, tracks the thought processes, decision-making intentions, and interaction rules of individual micro-intelligent agents, and identifies emerging cognitive patterns through clustering and temporal analysis.

[0019] The intent detection and clustering module extracts the intent sequences of individual agents in a multi-agent system through both bounded rationality and fully rationality channels in parallel; it embeds and clusters cross-agent intent representations, identifies emerging intent clusters, and calculates their relevance to behavior.

[0020] The active experiment planning module is based on the principle of maximizing information gain, iteratively selecting the next intervention variable and intervention intensity to form an active experiment planning closed loop of "experiment-learning-re-experiment".

[0021] The visualization and reporting module outputs metrics such as the performance and fairness of the multi-agent system, as well as causal explanation reports.

[0022] Furthermore, the observation and analysis unit uses conditional mutual information I(X;Y|Z) to perform variable edge selection and dependency structure learning, thereby obtaining a sparse candidate causal graph structure.

[0023] Furthermore, the intervention analysis unit uses random initial conditions and repeated simulations to calculate the average treatment effect and conditional treatment effect, and blocks the hybrid path through backdoor adjustment.

[0024] Furthermore, the mechanism analysis unit employs a structural causal algorithm to conduct an attribution-intervention-prediction counterfactual inference process: A) Attribution: performing a posteriori estimation of the exogenous noise distribution under observed facts; B) Intervention: performing a do operation on the target variable and running the model under sampled exogenous noise; C) Prediction: aggregating the output to obtain the empirical distribution and confidence interval of the counterfactual results.

[0025] Furthermore, the active experimental planning selects intervention variables and intensities according to maximizing information gain, and performs Bayesian updates on the posterior distribution of the causal graph.

[0026] Furthermore, the intervention analysis unit supports three types of operations: hard intervention, soft intervention, and uncertain intervention, and outputs causal effect estimates within a unified framework.

[0027] The present invention can also adopt the following technical solutions:

[0028] Construct an objective world model based on domain knowledge and historical data, and define the state, variables, events and transition rules of the multi-agent system.

[0029] Embed a subject-world model into each agent to enable it to make bounded rational decisions, learn adaptively, and evolve policies.

[0030] State trajectories are generated through multiple rounds of parallel simulation, and conditional mutual information, causal graph edge weights, outliers, and multi-trajectory probability distributions are calculated during the observation and analysis phase to identify potential causal structures and key perturbation points.

[0031] In both the objective world model and the subjective world model, do-interventions were implemented on key variables, and the average treatment effect and conditional treatment effect were calculated according to the following formulas;

[0032]

[0033]

[0034] In this model, X is the treatment variable, with a value of 1 indicating intervention and a value of 0 indicating no intervention; Y is the outcome variable; E[·] is the mathematical expectation operator; and do(X=x) represents the exogenous specification of variable X in the causal model. Therefore, ATE represents the average difference in outcomes between intervention and no intervention at the population level. Z is a covariate (such as individual attributes or environmental conditions), and Z = z represents the conditional expectation calculated under the condition of a value of z. CATE(z) then represents the conditional average treatment effect of outcome Y when intervention and no intervention are implemented in a population or situation given a covariate value of z.

[0035] By introducing monitoring of individual agents to extract their thought processes in real time, and using a large language model interface to obtain two types of intentions: bounded rationality and goal orientation, key emergent thoughts are identified and integrated into a group intention database. Common cognitive patterns are extracted through cosine similarity and k-means clustering to form an intention temporal emergence graph, thereby realizing the mapping of micro-cognitive mechanisms to macro-behavior.

[0036] Beneficial effects

[0037] This invention constructs an artificial social computing experimental platform that integrates objective world models and subjective world models, achieving multi-level identification and mechanistic explanation of causal relationships in complex systems, demonstrating significant technical and application advantages. This method overcomes the limitations of traditional causal inference, which can only identify variable correlations but struggles to reconstruct generative paths. It integrates observational analysis, intervention analysis, and mechanism analysis into a unified framework, enabling precise evaluation of variable intervention effects in simulated environments and in-depth tracking of the cognitive and behavioral evolution of micro-agents, thereby revealing the true causal mechanisms behind emergent phenomena in macro-systems. Simultaneously, this invention introduces a large language model to enhance agent intelligence, making intent recognition and behavioral clustering possible, effectively improving the explanatory power and operability of causal inference. This method is applicable to various complex system scenarios such as social governance, platform algorithm evaluation, and policy experiments, possessing strong versatility, scalability, and practical guiding value. Attached Figure Description

[0038] To better understand the method and its operating mechanism provided by this invention, the following figures are used to illustrate the structural composition and functional flow of this invention, and are only for illustrative purposes and not for limiting the invention:

[0039] Figure 1 : A structural diagram of the computational experimental causal analysis platform proposed in this invention;

[0040] Figure 2 The flowchart of the computational experimental causal analysis proposed in this invention;

[0041] Figure 3 Observation and analysis: A schematic diagram of multiple possible evolutionary trajectories in a complex system;

[0042] Figure 4 Intervention analysis: A causal modeling framework based on intervention;

[0043] Figure 5 Mechanism Analysis: Agent Intent Recognition Graph;

[0044] Figure 6 : Schematic diagram of the experimental application of this invention in managing involution in O2O platforms. Detailed Implementation

[0045] To more clearly illustrate the present invention, the following description is in conjunction with the accompanying drawings. Figure 1-4 and preferred embodiments Figure 5 The present invention will be further described below. Those skilled in the art should understand that the specific description below is illustrative and not restrictive, and should not be construed as limiting the scope of protection of the present invention.

[0046] Understanding causal mechanisms in complex systems requires both identifying the "causal effects" themselves and explaining the mechanisms behind the "causal processes." Computational experiments serve as a bridge connecting these two types of causal analysis. By simulating the operation of an "artificial society," computational experiments can not only predict causal effects between variables (positive causality) but also explore underlying causal mechanisms (reverse causality) by tracing simulated trajectories and counterfactual paths. This gives computational experiments a dual function in causal research: on the one hand, it provides algorithmic estimates of "causal effects"; on the other hand, it provides actionable modeling tools for "causal explanations."

[0047] This analytical capability relies on the construction and manipulation of world models, such as... Figure 1 As shown. A world model broadly refers to an abstract representation of the structure and evolutionary laws of a system in the real world. It can be divided into two categories: one is the external, objective world model (such as the Sora model), which constructs an accurate simulation of environmental state transitions through learning data, videos, and other information; the other is the internal, subjective model, namely the agent's belief model, used to express the agent's understanding and expectations of the world. Computational experiments, by embedding these two types of models, can not only simulate the evolutionary process of the external environment but also deduce the agent's behavioral decision-making paths under different cognitive premises.

[0048] Therefore, within the computational experimental analysis framework, forward causal analysis is typically achieved through intervention simulation: by changing the setting of a certain variable and observing the changes in the system output, a robust causal network graph between variables can be uncovered to describe the interactions within the system. Reverse causal analysis, on the other hand, relies on counterfactual trajectory generation: constructing paths that closely resemble actual trajectories but with variations in key variables, to infer potential causal paths or mechanisms. These two types of causal analysis work together during the simulation process to reveal the operational rules and behavioral evolution logic of the world model.

[0049] The analytical framework for computational experiments proposes analyzing experiments at three levels: the first level is association (what is?), i.e., given a variable, what is the probability of the outcome?; the second level is intervention (what if?), i.e., if something is done, what will the result be?; and the third level is mechanism (how is?), i.e., if the cause leads to the result, what is the process? The specific techniques and methods used at each level of the framework will be discussed below.

[0050] (1) Observational Analysis (Connecting Necessity and Contingency): This process is based on an objective world model and belongs to the positive causal research paradigm. The core viewpoint of observational analysis acknowledges the high degree of uncertainty in the future outcomes and paths of events, specifically manifested in the multiple possibilities of the evolution of social systems, which are influenced by initial conditions and external factors. Observational analysis focuses on modeling the possibility space of social system evolution paths, identifying all possible states that the social system may experience and their probability distribution. This process emphasizes the uncertainty of event evolution and reveals the randomness and uncertainty of social system evolution through scenario evolution data. When faced with multiple possible futures, observational analysis attempts to quantify each possible path and assess its probability of occurrence through a probabilistic framework. This approach not only reveals the contingency in the evolution of social systems but also provides theoretical support for dealing with different possibilities in the future.

[0051] like Figure 1 As shown, this invention proposes a computational experimental causal analysis system for complex systems, addressing three core bottlenecks in existing causal inference techniques for complex system scenarios: first, the causal mechanism is invisible, and traditional methods can only identify variable correlations but cannot explain the generation path; second, intervention is limited, as real-world experiments are difficult to conduct due to ethical, cost, or non-reproducibility issues; and third, there is a lack of unified modeling from micro to macro levels, making it difficult to simultaneously analyze causal effects and causal mechanisms within the same framework. Based on a multi-agent artificial social model, this invention integrates key technologies such as structural causal modeling (SCM), counterfactual simulation, world model building, large language modeling (LLM), intent analysis, and dynamic intervention to form a complete multi-level causal inference system. This system enables full-chain causal analysis and verification of complex systems in a virtual environment.

[0052] This invention uses computational experiments as its core technology. By constructing an artificial social model parallel to the real world, it can controllably simulate the evolution of a system under both natural and controlled states, thereby obtaining causal effect estimates and mechanism explanations without relying on real-world experiments. The innovation of this method lies in the simultaneous introduction of a dual-world model and a three-layer causal analysis framework. The dual-world model includes an "objective world model" and a "subjective world model." The objective world model expresses the application domain environment, structure, state transitions, and variable dependencies, serving as the operational vehicle for the entire experiment; the subjective world model is embedded in each agent, characterizing the individual's cognition, intentions, learning, and decision-making logic regarding the environment. Through this dual model, researchers can simultaneously observe the external state evolution and internal cognitive mechanisms of the system, capturing the complete path from micro-behavior to macro-emergence.

[0053] In the three-layer causal intervention module, the first layer, "Observational Analysis," identifies the state space, event sequence, and probability distribution of complex systems under natural evolution, constructing a potential causal graph structure to provide a baseline for intervention analysis. The second layer, "Intervention Analysis," applies exogenous interventions to key variables, conducts parallel simulations between the treatment and control groups, quantifies the true causal effect of the intervention on the system output, and automatically generates a structural causal model (SCM). The third layer, "Mechanism Analysis," delves into the agent's internal world model, tracks the thought processes, decision-making intentions, and interaction rules of micro-individuals, and identifies emerging cognitive patterns through clustering and temporal analysis, thereby revealing the true generative mechanism behind macroscopic phenomena. These three layers of analysis are nested and interconnected, and can operate independently or in combination, forming a closed loop from effect identification to mechanism explanation.

[0054] At the implementation level, this invention provides a standardized operating procedure: First, an objective world model is constructed based on domain knowledge and historical data, defining system states, variables, events, and transition rules; second, a subjective world model is embedded in each Agent, enabling it to possess bounded rationality decision-making ability, adaptive learning ability, and strategy evolution ability; then, state trajectories are generated through multi-round parallel simulation, and conditional mutual information, causal graph edge weights, outliers, and multi-trajectory probability distributions are calculated during the observation and analysis phase to identify potential causal structures and key perturbation points; in the intervention analysis phase, hard, soft, or uncertain interventions are applied to variables, parallel experiments are constructed, and the Pearl back-door criterion is used to estimate causal effects; finally, in the mechanism analysis phase, a monitoring Agent is introduced to extract the thought processes of each Agent in real time, and two types of intentions—bounded rationality and goal orientation—are obtained through a large language model interface. Key emergent thoughts are identified and integrated into a group intention database, and common cognitive patterns are extracted through cosine similarity and k-means clustering to form an intention time-series emergent graph, realizing the mapping from micro-cognitive mechanisms to macro-behavior. Figure 2 As shown.

[0055] Compared with existing technologies, this invention represents a substantial advancement in the following aspects: First, this method overcomes the limitation that lateral causal analysis can only verify effects but not explain processes, enabling integrated reasoning from "positive causality" to "reverse causality" within the same framework. Second, by reconstructing counterfactual trajectories in a virtual environment through computational experiments, it supports exploring the global sensitivity and causal path stability of complex systems under different variable combinations and initial conditions, significantly improving the robustness of causal identification. Third, by embedding the agent's world model and large language model interface, the simulated agent possesses realistic decision-making logic and cognitive evolution processes, thereby greatly enhancing the model's realism and explanatory power for human social behavior. Fourth, this invention provides scalable algorithmic processes and toolsets, supporting rapid deployment in complex systems across multiple domains, scales, and levels, including application scenarios such as social governance, economic policy evaluation, platform algorithm control, public safety early warning, environmental management, and ecosystem simulation.

[0056] In practical applications, the effectiveness of this invention has been verified through simulations of the "involution" phenomenon among O2O platform delivery riders. Experiments show that this method can identify the evolutionary trajectory of the degree of involution within a 30-day simulation window, extract the causal effects of intervention variables on system behavior, and track how key intentions such as anxiety and risk aversion gradually accumulate into macro-level competitive behavior. This case demonstrates the comprehensive ability of this invention to explain complex social phenomena, evaluate intervention strategies, and reveal behavioral mechanisms. Figure 3 A structural causal model (SCM) is often used to represent generative relationships between different variables in reinforcement learning, causal inference, or dynamic systems.

[0057] In complex social systems, there is a strong nonlinear coupling and dynamic feedback between individual behavior and environmental state. The evolutionary outcome of the system is usually the product of games, interactions, and adaptations among multiple agents, exhibiting high path dependence and multiple trajectory possibilities. Any observed event trajectory is merely one of many possible paths in the potential space, and its evolutionary structure is simultaneously influenced by initial conditions, background variables, and accidental perturbations. The core task of observational analysis is to construct the system's state evolution possibility structure, that is, to identify all possible state combinations, event sequences, and their probability distributions under natural operating conditions. This process is based on an "objective world model," which is usually derived from learning and modeling historical observation data. This model aims to abstract the system's state transition mechanisms, variable dependency structures, and evolutionary laws, providing a basic framework for understanding the system's natural evolution.

[0058] In this analysis, the operational trajectory of an artificial society can be formalized as a time-series trajectory of triples of state, event, and action. Based on this structure, causal relationships in the trajectory can be further defined: if event C leads to event E in the actual trajectory π, there may exist a counterfactual trajectory π' that is sufficiently similar to π, where the state of C is the same as in π, but the state of E is not the same as the state that occurred in π. This indicates that the same initial setting of an artificial society may generate various different counterfactual trajectories, which represent various traversal paths in the state space of the artificial society model. The simulated trajectory can be very complex, and depending on the occurrence and omission of events in trajectory π, several forms of causal relationships may arise: basic causal relationships, prevention, transitivity, conditional absence relationships, prevention propagation, etc.

[0059] The aforementioned trajectory structure not only helps identify causal paths between system variables, but also allows us to reconstruct the system's state evolution network and underlying mechanisms based on natural observation data without any intervention. To further formalize the strength of dependencies between variables, this invention introduces conditional mutual information to measure the statistical coupling relationship between variable triples.

[0060]

[0061] This index measures the degree to which variables X and Y share information given variable Z. If the mutual information is significantly non-zero, it can be considered that X and Y still have a dependent path after controlling for Z. In constructing models of the objective world, conditional mutual information can be used for variable edge selection, structure learning, and the construction of latent causal graphs, making it an important tool for observation and analysis.

[0062] To gain a deeper understanding of system evolutionary deviations, anomaly detection methods are generally employed to capture the random events that cause these deviations, including global and local anomaly detection. Machine learning techniques are used to identify complex deviations from expected behavior, while rule-based methods focus on detecting violations of predefined norms. In dynamic and high-dimensional environments, dimensionality reduction techniques, such as PCA and online learning algorithms, are needed to achieve adaptation. By combining causal discovery with anomaly detection, this framework not only reveals the potential evolutionary paths of the system but also explicitly identifies key deviations that may occur in its evolution.

[0063] Algorithm flow for observation and analysis of Algorithm 1.

[0064]

[0065] (2) Intervention Analysis (Connecting Causes and Consequences): This process is also based on an objective world model and belongs to the positive causal research paradigm. Intervention analysis uses randomized controlled experiments to compare the differences in the evolutionary trajectory of the system in different "hypothetical" scenarios, observe how changes in causal variables cause changes in outcome variables, and thus uncover robust structural causal models between variables. Intervention analysis mainly includes the following elements: experimental group / control group, sample consistency, random sampling / random grouping, etc. By running the control group and experimental group in parallel, the average experimental effect (ATE) can be used to quantify the impact of the intervention on the system. Figure 4 As shown, this describes a concept of causal analysis based on a network or graph model. It focuses more on the conceptual framework, illustrating the logical chain from "intervention" to "outcome." A specific structural equation model is described, serving as the mathematical tool for realizing the conceptual framework on the left. It demonstrates the concrete quantitative relationships between variables. The four elliptical nodes on the left side of the conceptual framework for causal intervention and their relationships elucidate the core idea of ​​causal inference.

[0066] In traditional social science research, conducting randomized controlled trials often faces ethical, cost, and reproducibility obstacles. In contrast, computational experiments provide a natural counterfactual platform for causal inference. By constructing and manipulating virtual worlds, researchers can create two "parallel worlds" within a unified system model: one undergoing intervention, and the other remaining in its natural evolution, thus allowing for direct comparison of treatment effects without confounding bias. Furthermore, the identifiability of the intervention effects in computational experiments is guaranteed by Pearl's backdoor adjustment formula.

[0067] The core of intervention analysis lies in simulating and comparing the evolutionary differences of a system under "with intervention" and "without intervention" paths. Using a unified world model, the trajectories of the treatment and control groups can be constructed within the model, thereby obtaining counterfactual estimates of causal effects. This invention expresses the average treatment effect (ATE) under intervention in the following form:

[0068] (2)

[0069] Here, do(X = x) represents an exogenous specification of variable X, "severing" all causal paths pointing to X from the system mechanism, retaining only the direct or indirect causal effect of X on Y. To ensure the identifiability of the estimate, Pearl's backdoor adjustment criterion must be satisfied, i.e., there exists a set of variables Z that can block all non-causal paths between X and Y. When dealing with binary variables, x = 1, x′ = 0 is generally taken.

[0070] In simulation systems, we typically select Z using domain knowledge, sensitivity analysis, or structural learning algorithms, and calculate the intervention effect using the following formula:

[0071] (3)

[0072] Z blocks all non-causal paths between X and Y, thus ensuring the identifiability of the intervention effect estimate. By obtaining empirical estimates of the above terms through repeated experiments, the true impact of the intervention variable on the system's evolution can be effectively identified.

[0073] Intervention is achieved by modifying the values ​​of a subset of variables in a model. Generally, there are several types of intervention: hard intervention, soft intervention, uncertain intervention, and no intervention. Taking epidemic transmission as an example, intervention analysis can be used to identify transmission chains, estimate the blocking effect of intervention measures, and even locate "patient zero." In practice, a counterfactual simulation without intervention is first conducted, setting only the initial infection state and allowing the system to evolve naturally. Subsequently, a series of active intervention experiments are designed, such as modifying variables like the transmission coefficient, incubation period, and network topology. Each experiment fixes one variable, and the distribution of results is obtained through simulation, updating the causal graph structure until convergence to stability or the preset experimental termination conditions are met.

[0074] Table 2 Algorithm Flow of Intervention Analysis

[0075]

[0076] (3) Mechanism Analysis (Connecting Micro and Macro): This process follows a reverse causal research paradigm, focusing not only on the objective world model but also on analyzing the agent's subjective world model. Y represents the social phenomenon to be explained, X represents several subjects with different attributes, and rule represents the agent's world model, revealing the agent's internal cognitive structure and thought process during decision-making, game theory, or action. Mechanism analysis argues that the actions taken by the agent and the agent's interactions are the causes of macro phenomena, while mechanistic explanations show how the agent's actions and interactions lead to specific phenomena "step by step," such as... Figure 5 As shown.

[0077] Mechanism analysis focuses on the construction of an individual agent's internal world model and the explanation of behavioral decision-making mechanisms. Starting from the agent's micro-cognitive structure (such as beliefs and intentions), it characterizes how agent interaction behaviors emerge as system-level changes. We first formalize the mechanism as connecting causal variables. With outcome variable Intermediary structure ,

[0078] (4)

[0079] in, This represents a causal mechanism with an internal process structure. To characterize the internal evolution of the mechanism, we further expand it into a state sequence path:

[0080] (5)

[0081] Here, Represents a series of intermediate states in a causal chain, describing the progression from the initial cause. result The multi-step evolutionary path. This unfolding structure illustrates that the macroscopic result Y is generated by the gradual accumulation of a sequence of microscopic states.

[0082] However, in complex systems, these state evolutions often do not occur statically, but rather unfold dynamically over time. Therefore, we introduce a temporal representation to model the above sequence as a set of causal rules. Evolutionary trajectory:

[0083] (6)

[0084] here, Indicates the initial conditions of the system; Indicates the first The causal rules of each step (e.g., agent response, game theory, cognition, etc.); This is a system function that aggregates the entire trajectory into an output result. Through this form, we transform the static path structure into a mechanism expression with dynamic generation capabilities. So these... Where does it come from? Let's go back to the structural definition of the mechanism:

[0085] To fully represent the components involved in the mechanism structure, we define the final mechanism form as a triple:

[0086] (7)

[0087] in, It is the event space (system state, agent behavior, etc.); It is a space of causal rules (response functions, cognitive models, etc.); It represents a causal transfer relationship, indicating how events transition under rule-driven conditions. This ternary structure can support micro-level cognitive modeling (such as Beliefs, Desires, and Intentions) as well as rule enforcement and multi-agent interactions, thereby enabling the full-process modeling and simulation of causal paths. regular sequence It is precisely from mechanism R that the elements are selected and applied to the event space in the prescribed order. .

[0088] The above representation can support both micro-level cognitive modeling (such as Beliefs–Desires–Intentions) and rule execution and multi-agent interaction, thereby enabling full-process modeling and simulation of causal paths. The generated rule sequence is the result of selecting rules from the rule space R in a predetermined order and applying them sequentially to the event space E. Therefore, mechanism analysis not only defines "what it is" (structure) but also specifies "how it is generated" (process), ultimately deriving the evolutionary outcome of the system.

[0089] Mechanism analysis reinforces the dual modeling path of the "world model": ① On the one hand, it relies on the "objective world model" to describe the evolutionary laws of macroscopic phenomena. ② On the other hand, it starts from the agent's "internal world model," generates microscopic behavioral trajectories, and identifies system emergence through cluster analysis.

[0090] Table 3. Algorithm flow for mechanism analysis

[0091]

[0092] Through the above three-layer causal analysis structure, the computational experimental method proposed in this invention provides systematic support for understanding the behavior of complex systems, reconstructing causal chains, and assisting in decision-making and intervention, opening up new avenues for causal analysis to move from quantitative prediction to mechanism explanation.

[0093] To verify the effectiveness of the method in complex real-world systems, this paper takes O2O instant delivery platforms as the research object and conducts computational experiments focusing on the real-world social phenomenon of "rider involution." In O2O platform operations, riders constantly improve their order-taking efficiency and service timeliness to obtain more orders. This competitive mechanism evolves to some extent into disordered involution, potentially leading to systemic problems such as income fluctuations, behavioral alienation, and decreased platform efficiency. Therefore, this invention constructs an artificial social simulation system to reproduce the entire process of order taking, order bidding, delivery, and feedback on the platform. Through multiple rounds of experiments, the causal structure between intervention variables and system outputs is analyzed, and the potential mechanism paths that trigger involution are further identified.

[0094] like Figure 6As shown, the experimental platform is designed as follows: The system includes three main entities: the platform, riders, and users. The platform is responsible for order generation and dispatch allocation. Users submit order requests, and riders make order acceptance decisions, route planning, and fulfillment execution based on algorithms or experience. The platform environment includes parameters such as geospatial, temporal dimensions, and order hotspots, and considers uncontrollable disturbance variables in reality, such as traffic congestion, weather fluctuations, and customer reviews. Riders, as agents, possess cognitive models and can dynamically respond based on historical experience, rules, or learning strategies. Table 1 shows the mapping relationship between specific delivery scenarios and the computational experimental system.

[0095] The experiment employed multiple parallel simulations and control groups to compare the evolution of system behavior under different intervention measures (such as order dispatch strategies, incentive schemes, and information transparency). Intervention variables included, but were not limited to, rider incentive intensity, platform algorithm bias, task transparency, and order distance distribution. Output indicators covered rider average order acceptance time, task allocation fairness, rider earnings differences, and changes in rider workload.

[0096] Table 4. Mapping relationship between the real world and computational experimental systems

[0097]

[0098] During the experimental execution phase, this invention constructs a unified world model, sets the initial system state, and conducts parallel experiments using the following process:

[0099] 1) Perform multiple rounds of simulation for each set of variables and collect experimental output.

[0100] 2) The causal graph structure learning method is used to construct the dependency paths between variables and to initially establish the structural causal graph.

[0101] 3) Use intervention analysis to apply the do operation to a single variable, compare the differences in output before and after the variable change, and verify the path directionality and causal strength.

[0102] 4) Introduce a counterfactual trajectory comparison mechanism to generate multiple nearest neighbor trajectory samples and screen for key variable nodes and path redundancy.

[0103] 5) Employ mechanism analysis methods, combining the agent's internal belief update process and intention evolution trajectory, to identify systemic behavioral biases caused by micro-rules.

[0104] Experimental results show that platform information transparency has a significant guiding effect on riders' order-accepting strategies, and high transparency can effectively alleviate the anxiety and behavioral distortions caused by competition for orders. Compared with high-intensity performance incentive strategies, stable incentives and clear task boundaries are more conducive to creating a healthy competitive environment. Further follow-up analysis shows that the key mechanism pathways of involution behavior mainly include: cognitive misalignment → high expectations → excessive effort → pressure accumulation → intensified game → behavioral alienation → system fluctuation.

[0105] Through the above experiments, the computational experimental causal analysis framework proposed in this invention can not only deeply characterize the involution mechanism of O2O platforms, but also provide decision support for the design and optimization of intervention policies, providing a general path for dealing with evolutionary problems in other complex social systems.

[0106] It should be noted that, in order to enhance the adaptability and versatility of the experimental design, the application of the method proposed in this invention to the "rider involution" problem is highly similar to causal reasoning tasks in other complex system contexts. The causal modeling and computational experimental strategies provided in this invention can also be adapted, and their core mechanism can be compared to the collaborative evolution process of a multi-agent system. The relevant principles and analysis procedures can be referred to the aforementioned computational experimental framework and causal reasoning method, and will not be repeated here.

[0107] Some terms used in this invention specification, such as "input," "output," "module," and "relationship between variables," are used to express the abstract functional division and causal structural relationships in the process of complex system modeling, and should not be narrowly interpreted as functional components of hardware devices or physical structures. These terms are used to describe the functional roles of the computational experimental system in processes such as logic modeling, simulation execution, and intervention analysis, and the specific implementation may vary depending on different simulation platforms, modeling tools, or application fields.

[0108] Furthermore, unless the context explicitly specifies otherwise, terms used in this specification such as "establish," "connect," "embed," "generate," and "evolve" should be understood from the perspective of model abstraction and system behavior. They can represent functional mappings between variables at the conceptual level, dependencies in the model structure, dynamic interactions between the agent and the environment during simulation, or nonlinear transitions of system-level behavior over time. Therefore, the use of these terms does not imply mandatory limitations on the implementation methods of system components, but rather serves to express the causal logical relationships of this invention at the theoretical construction and model implementation levels.

[0109] The designations such as "first causal layer" and "second causal layer" mentioned in this invention are intended to distinguish different levels of causal analysis logic and modeling dimensions, and do not imply a specific execution order or model constraint level. They are used to characterize the complete analysis path from correlation identification and causal intervention to mechanism explanation in complex systems, forming interdependent yet separable modeling layers within the computational experimental framework.

[0110] It should be noted that terms such as "including" and "comprising" are used in this invention to characterize the openness and extensibility of the methodological framework, expressing a non-exclusive, inclusive structure. For example, when constructing a causal graph, although this invention clarifies the identification of key variables and the path construction process, it does not preclude the introduction of additional mediating variables, confounding factors, or specific rule modules depending on the application scenario. Therefore, any algorithm extensions, module replacements, or scenario transfer applications derived from the technical ideas of this invention, as long as they do not depart from the core principles of this invention, should be considered within the reasonable scope of protection of this invention.

[0111] Obviously, the embodiments described in this invention are merely one typical application illustrating its core methodology, aiming to more clearly present the integration of computational experiments and causal reasoning. For those skilled in the art, various modifications, substitutions, or combinations can be made without departing from the inventive concept and causal framework, and such improvements should also be included within the scope of protection of this invention.

Claims

1. A platform for multi-agent system causal effect analysis, characterized in that: The platform comprises a data processing module, an objective world model, a subject world model, a causal intervention module, an intention detection and clustering module, an active experiment planning module, and a visualization and reporting module; the causal intervention module comprises an observation analysis unit, an intervention analysis unit, and a mechanism analysis unit; wherein The data processing module abstracts the real world into relevant data of an artificial society or field, and obtains panel data through computational experiments; The objective world model is used to express the application field environment, structure, state transition, and variable dependence, and is the state transition, variable dependence, and evolution law of the entire experiment under the observation data driving; The subject world model is used to be embedded into each intelligent agent, and is used to depict the individual's cognition, intention, learning, and decision logic for the environment; The observation analysis unit is used to identify the state space, event sequence, and probability distribution of a complex application field under natural evolution, and to construct a potential causal graph structure; The intervention analysis unit quantifies the real causal effect of an intervention on the output of an application field by applying an exogenous intervention to key variables and performing parallel simulation between a treatment group and a control group; The mechanism analysis unit tracks the thinking path, decision intention, and interaction rule of a micro intelligent agent individual by deeply analyzing the world model inside a single intelligent agent, and identifies emerging cognitive patterns through clustering and time series analysis; The intention detection and clustering module extracts individual intelligent agent intention sequences through a bounded rationality channel and a complete rationality channel in parallel for a multi-agent system; embeds and clusters cross-agent intention representations, identifies emerging intention clusters, and calculates the correlation with behaviors; The active experiment planning module iteratively selects the next intervention variable and intervention intensity based on the principle of maximum information gain, forming an active experiment planning closed loop of "experiment-learning-reexperiment"; The visualization and reporting module outputs multi-agent system efficiency and fairness indicators as well as causal explanation reports.

2. The platform for multi-agent system causal effect analysis of claim 1, wherein: The observation analysis unit uses conditional mutual information I(X;Y|Z) for variable edge selection and dependence structure learning to obtain a sparse candidate causal graph structure.

3. The platform for multi-agent system causal effect analysis of claim 1, wherein: The intervention analysis unit uses random initial conditions and repeated simulation to calculate the average treatment effect and the conditional treatment effect, and blocks the confounding paths through backdoor adjustment.

4. The platform for causal effect analysis of a multi-agent system according to claim 1, characterized in that: The mechanism analysis unit uses a structural causal algorithm to carry out a cause-intervention-predict counterfactual inference process: A) Cause: posteriori estimation of exogenous noise distribution under observed facts; B) Intervention: do operation on target variables and run the model under sampled exogenous noise; C) Prediction: aggregate output to obtain the empirical distribution and confidence interval of counterfactual results.

5. The platform for causal effect analysis of multi-agent systems of claim 1, wherein: The active experiment planning selects intervention variables and intensity according to the principle of maximum information gain, and performs Bayesian update on the posterior distribution of the causal graph.

6. The platform for causal effect analysis of multi-agent systems of claim 1, wherein: The intervention analysis unit supports three types of operations: hard intervention, soft intervention, and uncertain intervention, and outputs causal effect estimates in a unified framework.

7. A method for analyzing a multi-agent system based on causal effects, the method comprising: The method is implemented based on the platform as claimed in claims 1-6, comprising the following steps: According to the field knowledge and historical data, an objective world model is constructed to define the state, variable, event and transition rule of multi-agent system; A subject world model is embedded in each agent to make it have the ability of limited rational decision, self-adaptive learning and strategy evolution; Through multiple rounds of parallel simulation, state trajectories are generated, and conditional mutual information, causal graph edge weight, abnormal points and multi-trajectory probability distribution are calculated in the observation and analysis stage to identify potential causal structure and key disturbance points; In the objective world model and the subject world model, do intervention is implemented on the key variables, and the average treatment effect and the conditional treatment effect are calculated according to the following formula: ATE = E[Y | do(X=1)] − E[Y | do(X=0)]; CATE(z) = E[Y | do(X=1), Z=z] − E[Y | do(X=0), Z=z]; Where X is the treatment variable, taking the value 1 means implementing intervention, and taking the value 0 means not implementing intervention; Y is the result variable, E[·] is the mathematical expectation operator, and do(X=x) means external setting of variable X in the causal model. Thus, ATE represents the average result difference between implementing intervention and not implementing intervention at the population level; Z is the covariate (such as individual attributes or environmental conditions), Z=z means calculating the conditional expectation under the condition that the value is z; CATE(z) represents the conditional average treatment effect of result Y when implementing intervention and not implementing intervention in the population or situation with given covariate value z; A monitoring module is introduced to monitor individual agents in real time and extract their thinking streams. With the help of large language model interface, two types of intentions, limited rationality and goal orientation, are obtained, key emerging thoughts are identified and integrated into the group intention library, common cognitive patterns are extracted through cosine similarity and k-means clustering, and intention time sequence emergence diagram is formed to realize the mapping from micro cognitive mechanism to macro behavior.

Images