Performance evaluation method for time sequence cognitive attributes of multi-agent system

Abstract

The invention discloses a performance evaluation method for time sequence cognitive attributes of a multi-agent system, and aims to solve the problems of inaccurate and incomplete performance evaluation of the multi-agent system in the prior art. According to the method, firstly, a concurrent game structure is adopted to model a multi-agent system, and key performance attributes of the multi-agent system are described and evaluated by introducing a performance evaluation language combining sequential logic and cognitive logic. While the time sequence characteristics are expressed, the knowledge relationship between the intelligent agents can be captured, so that more comprehensive performance evaluation is provided. According to the method, a recursive calculation nested formula mode is adopted, the atomic formula value of the innermost layer is calculated firstly, then calculation is conducted outwards layer by layer, and finally the numerical value of the performance attribute formula of the whole system is obtained. The method can accurately evaluate the performance of the multi-agent system, including but not limited to key indexes such as reaction time, safety, cooperation efficiency and energy consumption. According to the method, reference can be provided for design and optimization of the multi-agent system, and a quantifiable basis can be provided for task scheduling, resource allocation and strategy selection.

Description

Technical Field

[0001] The present invention discloses a performance evaluation method for the temporal cognitive attributes of a multi-agent system. The method combines temporal logic and cognitive operators and can quantitatively analyze key performance indicators such as the collaborative behavior, reaction time, and safety of the multi-agent system in a complex interactive environment. The method is suitable for multi-agent systems such as autonomous driving fleets. Background Art

[0002] Multi-agent systems (MAS) consist of multiple agents, each of which can independently perceive the environment, make decisions, and perform actions. These agents collaborate to complete complex tasks and are widely used in autonomous driving, drone swarms, smart manufacturing, and other fields. Due to the high degree of collaboration and dynamic interaction between agents in the system, how to accurately evaluate the performance of multi-agent systems has become a research focus. Existing performance evaluation methods face many challenges in dealing with these complexities and dynamics.

[0003] Existing evaluation languages, such as sequential logic CTL and PCTL, are widely used in the performance evaluation of dynamic systems. CTL can describe the changes in states under different paths in the system through a tree structure, and is particularly suitable for the analysis of concurrent and parallel systems. However, although these languages ​​have strong expressive power in sequential logic, they cannot fully express the knowledge sharing and cognitive process between agents in multi-agent systems.

[0004] Cognitive logic is a logical form that can express knowledge sharing and reasoning capabilities among agents. By introducing cognitive operators K, common knowledge E, and public knowledge C, cognitive logic can describe an agent's cognition and reasoning of a specific state, as well as its inference of other agents' cognition. This enables cognitive logic to capture complex knowledge interactions and agent collaboration in multi-agent systems.

[0005] Based on knowledge operators, this paper proposes a multi-agent system performance evaluation method that combines temporal logic and cognitive logic. This method can not only express the cognition of the agent, but also better express the interaction and collaboration between agents. This method that combines cognitive logic with dynamic temporal logic can be more comprehensive and accurate in the performance evaluation of multi-agent systems. Summary of the invention

[0006] The present invention proposes a performance evaluation method based on Multi-Agent Performance Evaluation Language (MAEL). This method combines temporal logic and cognitive logic on the basis of concurrent game structure, and realizes real-valued performance evaluation of multi-agent systems by recursively calculating performance attribute formulas.

[0007] The present invention is a performance evaluation method for the temporal cognitive attributes of a multi-agent system, which mainly comprises the following steps:

[0008] Step 1: Use the concurrent game structure to model the multi-agent system, describe the behavior and decision-making of the agents in different states, and characterize the system as a tuple G = (S, S0, M, A, Ac, AF, δ, L, F a(a∈A) );

[0009] Step 2: Use the Multi-Agent Evaluation Language (MAEL) to formulate the system performance attributes. The formula includes atomic propositions, operators (+, ×, ∧, ∨, MIN, MAX), timing operators (U, X) and knowledge operators (K, E, C);

[0010] Step 3: Provide a specific performance evaluation algorithm that can recursively calculate the performance formula layer by layer, starting from the innermost atomic formula to the outer formula, to obtain the final performance evaluation value of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 Flowchart of the performance evaluation method for the temporal cognitive attributes of multi-agent systems

[0012] Figure 2 Flowchart of performance evaluation algorithm

[0013] Figure 3 The concurrent game structure diagram for the driverless car fleet (in the case description section) DETAILED DESCRIPTION

[0014] The implementation of the present invention provides a performance evaluation method for the temporal cognitive attributes of a multi-agent system. In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. The implementation methods described with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.

[0015] The purpose of the present invention is to solve the problem that the existing technology cannot effectively evaluate the complex interactions and collaborative behaviors between agents. By combining temporal logic and cognitive logic, the present invention can accurately evaluate key performance indicators in multi-agent systems, such as response time, system security, etc. The implementation method of the present invention focuses on system modeling, definition of performance attribute formulas and implementation of recursive calculations, which can effectively deal with the uncertainty and dynamics in multi-agent systems and provide a scientific basis for system design and optimization.

[0016] The flowchart of the performance evaluation method for the temporal cognitive attributes of multi-agent systems is shown in the figure. Figure 1 The specific implementation method of the invention is as follows:

[0017] 1. Modeling multi-agent systems using concurrent game structures

[0018] The concurrent game structure is a model that can describe multiple agents taking independent actions at the same time and jointly affecting the system state. The thinking state is added to the existing concurrent game structure. CGS is a tuple, G = (S, S0, M, A, Ac, AF, δ, L, F a(a∈A) ),in

[0019] (1) S represents a set of states;

[0020] (2) S0 represents the set of initial states;

[0021] (3) M represents the set of mental states;

[0022] (4) A represents the set of agents;

[0023] (5) Ac represents a set of actions;

[0024] (6) AF represents a set of atomic formulas;

[0025] (7)δ: represents the transition relation, which defines the transition of states based on the joint action g. If (s, g, s′)∈δ, then s′ is the next state;

[0026] (8) is a function that labels each state with the value in that state, is the set of real numbers.

[0027] (9)F a represents the mental state extracted from the current state s, m a is the current state of mind, F a (s) = m a , where a∈A.

[0028] The present invention adopts appropriate data structures to represent the characteristics and operation rules of each part of the concurrent game structure, wherein the state set S, the initial state set S0, the thinking state set M, the agent set A and the atomic formula set AF are all stored in sets, so as to efficiently represent and operate these discrete element sets. a The dictionary is used to represent it, where the action set maps the agent to its action set through the dictionary, the state transition function maps the state and joint action to the next state, the labeling function maps the state and atomic formula to real values, and the thinking state function is used to record the thinking state of each agent in different states. Through the reasonable combination of the set and the dictionary, the present invention realizes the efficient modeling of the concurrent game structure. The specific structure is shown as follows:

[0029]

[0030] 2. Definition of performance attribute formula

[0031] The Multi-Agent Evaluation Language (MAEL) is an extension of CTML. It combines the advantages of CTL and PCTL, allowing rich temporal features that cannot be expressed in traditional discrete temporal logic. However, in multi-agent systems, since the understanding between agents must be considered, we introduce knowledge operators.

[0032] The syntax of MAEL is a temporal logic formula as follows:

[0033]

[0034] In addition, f∈AF,⊙ is a set of operators, including {+,×,∧,∨},K a is the knowledge operator of agent a, and a∈A, where A is the set of agents, and for all agents there exists a strategy that enables them to cooperate to achieve the goal, C A Denotes φ is the public knowledge in A, E A It means that each agent in A knows φ, and t represents the step.

[0035] For a given CGS G = (S, S0, M, A, Ac, AF, δ, L, F a(a∈A) ), s∈S is a state, [[φ]](s) is the mapping of the performance evaluation value of the current formula, and its value range is (-∞,∞). The semantics of MAEL can be defined as follows:

[0036] 1) [[f]](s) = L(s, f).

[0037] 2)[[φ1+φ2]](s)=[[φ1]](s)+[[φ2]](s).

[0038] 3)[[φ1×φ2]](s)=[[φ1]](s)×[[φ2]](s).

[0039] 4)[[φ1∧φ2]](s)=min{[[φ1]](s), [[φ2]](s)}.

[0040] 5)[[φ1∨φ2]](s)=max{[[φ1]](s), [[φ2]](s)}.

[0041] 6)

[0042] The value is 1 if the agent a knows φ, and the value is 0 if a does not know φ.

[0043] 7)

[0044] 8)

[0045] Where R a ={(s, s′)|F a (s′)=F a (s)), R A =U a∈A R a , YesR A The transitive closure of .

[0046] 9)[[MIN A Xφ]](s)=min{φ(s′)|δ(s, g)=s′}.

[0047] 10)[[MAX A Xφ]](s)=max{φ(s′)|δ(s, g)=s′}.

[0048] 11)

[0049] where ⊙∈{+,×,∧,∨}, represents the set of paths starting from state s in alliance A, where one agent adopts strategy π and the other agents choose randomly.

[0050] 12)

[0051] where ⊙∈{+,×,∧,∨}, represents the set of paths starting from state s in alliance A, where one agent adopts strategy π and the other agents choose randomly.

[0052] First, we define atomic formulas f. Each atomic formula corresponds to a basic state in the system. We use temporal logic operators to construct complex performance attribute formulas. Common temporal logic operators include: "next step" (X) and "until" (U). The specific temporal logic formula expression form is Xφ, φ1Uφ2. On this basis, we introduce knowledge operators K, K a φ is used to describe the knowledge state of agent a about formula φ. Introduce knowledge operator E, E A φ represents the knowledge state that all agents in the system know φ. Introducing knowledge operators C, C Aφ represents the public knowledge in the system, which means that all agents know and all agents know that other agents also know the knowledge state of φ. The performance attribute formula is constructed by combining the atomic formula, the timing operator and the knowledge operator, and the performance attribute formula is obtained by combining the operators addition (+), multiplication (×), conjunction (∧), disjunction (∨), minimum (MIN), and maximum (MAX).

[0053] In order to parse and calculate the performance attribute formula of the multi-agent system, the present invention proposes a data structure based on a syntax tree. The syntax tree is recursively constructed by decomposing the formula into basic operators and atomic formula nodes to represent the logical structure of the formula layer by layer. Each node of the syntax tree corresponds to an element in the formula, the root node represents the highest level operator of the formula, and the leaf node represents the most basic atomic formula.

[0054] For an atomic formula f, the atomic formula node is a leaf node; for a formula containing an operator ⊙, the root node of the syntax tree is ⊙, and the left and right child nodes represent the subformulas φ1 and φ2 on both sides of the operator, respectively; for the formula MIN A Xφ, the root node of the syntax tree is MIN A X and takes the syntax tree of φ as a child node; for the formula MAX A Xφ, the root node of the syntax tree is MAX A And take the syntax tree of φ as a child node; for the formula The root node of the syntax tree is And the syntax trees of φ1 and φ2 are taken as left and right child nodes; for the formula The root node of the syntax tree is And take the syntax trees of φ1 and φ2 as left and right child nodes; for formula K a φ, the root node of the syntax tree is K a And take the syntax tree of φ as the child node; for the formula K a φ, the root node of the syntax tree is K a And take the syntax tree of φ as a child node; for formula E A φ, the root node of the syntax tree is C A And take the syntax tree of φ as a child node. The specific structure is as follows:

[0055]

[0056] 3. Propose a performance evaluation algorithm to recursively calculate the evaluation results

[0057] This paper proposes a multi-agent system performance evaluation algorithm based on recursive calculation. The algorithm combines the multi-agent performance evaluation language (MAEL) and the concurrent game structure, and evaluates the performance indicators of the system through recursive calculation of nested formulas. Specifically, the algorithm starts with the calculation of atomic formulas, and gradually recursively processes operators, temporal logic formulas, and knowledge logic formulas, and finally obtains the overall performance evaluation result of the system. The flowchart of the algorithm is shown in the figure. Figure 2 shown.

[0058] 3.1 Input the concurrent game structure model CGS of the multi-agent system, the performance attribute formula φ and the current state s, and recursively calculate Comput(CGS, φ, s);

[0059] 3.2 When φ is an atomic formula f, the result returns the value of the atomic formula in the current state s;

[0060] 3.3 When φ=φ1+φ2, the result is returned as the sum of Comput(CGS,φ1,s) and Comput(CGS,φ2,s).

[0061] 3.4 When φ=φ1×φ2, the result is returned as the multiplication of Comput(CGS,φ1,s) and Comput(CGS,φ2,s).

[0062] 3.5 When φ=φ1∧φ2, the result returns the minimum value between Comput(CGS,φ1,s) and Comput(CGS,φ2,s).

[0063] 3.6 When φ=φ1∨φ2, the result returns the maximum value of Comput(CGS,φ1,s) and Comput(CGS,φ2,s).

[0064] 3.7 When φ includes K a When φ is calculated, the call Comput(CGS,K a φ, s), calculate the value of s′ in any state Comput(CGS,φ, s ′ ) is equal to Comput(CGS,φ,s). If they are equal, it means that the knowledge of agent a is the same, indicating that a knows that φ holds. The result returns 1, otherwise it returns 0 and records the corresponding label.

[0065] 3.8 When φ contains E A When φ is calculated, the call Comput(CGS,E A φ, s), traverse each agent a, if they all satisfy Comput(CGS,K a If the value of φ, s) is 1, it means that every agent knows that φ holds true, and the result returns 1, otherwise it returns 0 and records the corresponding label.

[0066] 3.9 When φ contains C A φ, the calculation calls Comput(CGS, C A φ, s), first initialize the transitive closure For any binary relation (s, s′) belonging to the transitive closure, if the value of Comput(CGS, φ, s′) is 1, it means that the knowledge is public knowledge and the result is returned as 1, otherwise it returns 0 and records the corresponding label.

[0067] 3.10 When φ contains MIN A When Xφ, the calculation calls Comput(CGS,MIN A Xφ, s), initialize the minimum value minValue = ∞, traverse the state set, if s′ is the next state of s under the joint action g, calculate Comput(CGS,φ, s′) and compare it with the minimum value and then update the minimum value minValue. The result returns the minimum value minValue and records the corresponding label.

[0068] 3.11 When φ contains MAX A When Xφ, the calculation calls Comput(CGS,MAX A Xφ, s), initialize the maximum value maxValue = -∞, traverse the state set, if s′ is the next state of s under the joint action g, calculate Comput(CGS,φ, s′) and compare it with the maximum value and then update the maximum value maxValue. The result returns the maximum value maxValue and records the corresponding label.

[0069] 3.12 When φ contains When the calculation calls Initialize the minimum value minValue = ∞, select the next state set that can be transferred to under the current state s according to the strategy, traverse the state set, and determine whether it is established from φ1 to φ2 when t steps are satisfied. Calculate the value of each state and accumulate or multiply it, traverse all paths, compare the calculation result with the minimum value, and update the minimum value minValue. The result returns the minimum value minValue and records the corresponding label.

[0070] 3.13 When φ contains When the calculation calls Initialize the maximum value maxValue = -∞, select the next state set that can be transferred to under the current state s according to the strategy, traverse the state set, and determine whether it is established from φ1 to φ2 when t steps are satisfied. Calculate the value of each state and accumulate or multiply it, traverse all paths, compare the calculation result with the maximum value, and update the maximum value maxValue. The result returns the maximum value maxValue and records the corresponding label.

[0071] The above calculation process is shown in Algorithm 1-8:

[0072]

[0073]

[0074] Implementation Cases

[0075] In order to verify the performance evaluation method of the temporal cognitive attributes of the multi-agent system proposed in this invention, we selected a scenario of an autonomous driving fleet driving in a complex urban environment for experimental analysis. This scenario is representative of real traffic and involves complex interactive behaviors of multiple agents (autonomous driving vehicles). By using the performance evaluation algorithm of this invention, we conducted a comprehensive evaluation of the fleet's key performance indicators of time and safety.

[0076] We use concurrent game structures to model the interactions between multiple autonomous vehicles, such as Figure 3 As shown. Each vehicle is regarded as an independent intelligent agent that can make autonomous decisions based on environmental perception information. The CGS model can accurately describe the interactive behaviors, cooperative relationships and perception of the environment between multiple intelligent agents. The modeled scenarios include situations where two vehicles are traveling in the same direction and encounter obstacles. Each state corresponds to the thinking state of the agent, such as the initial state m1, the obstacle detection state m2, the obstacle avoidance state m3, and the normal driving state m4. The available actions of the vehicle include deceleration, acceleration and obstacle avoidance. The state transition function δ describes the change of the system from one state to another under different action combinations. For example, when both vehicles are driving normally, the system transfers from the initial state s1[m1] to the state s2[m2] where the obstacle is detected.

[0077] We have conducted quantitative evaluations on multiple key performance attributes of the fleet. By substituting the performance attribute formula into the algorithm, we can calculate the performance evaluation results. The performance attribute formula and calculation results are defined as follows:

[0078] (1) Evaluate whether vehicle a knows the maximum time the convoy can maintain normal driving status,

[0079]

[0080] The calculation result is 1.

[0081] (2) After detecting an obstacle, evaluate whether all vehicles in the fleet know that they need to avoid the obstacle in the next state,

[0082] [[C V detected→MAX VXavoidObstacle]]

[0083] The calculation result is 1.

[0084] (3) Evaluate the maximum time required for all vehicles in the convoy to detect an obstacle, avoid it within 10 time units, and resume normal driving until all vehicles are aware of the obstacle.

[0085]

[0086] The calculation result is 1.

[0087] (4) Evaluate the minimum time required from obstacle detection to vehicle execution of obstacle avoidance maneuvers within 10 time units.

[0088]

[0089] The result of the calculation is 2.

[0090] (5) The longest continuous time during which all vehicles in the queue are running normally,

[0091]

[0092] The result is 10.

Claims

1. A performance evaluation method for temporal cognitive attributes of a multi-agent system, characterized in that: The main steps include: (1) Modeling multi-agent systems using concurrent game structures; (2) Use multi-agent performance evaluation language to define the performance attribute formula of the system; (3) Obtain the system performance evaluation results by recursively calculating the nested performance attribute formula.

2. The method for evaluating the performance of temporal cognitive attributes of a multi-agent system according to claim 1, characterized in that: The described step (1) is specifically: (1) First, define the set S of all possible states in the multi-agent system. Each state s∈S represents the specific state of the system at a certain moment. Define the initial state set S0∈S, which represents the possible states of the system at the beginning. (2) Define a set A of agents participating in the system operation. For each agent a∈A, define an action set Ac, which defines the actions that the agent can take in a given state. (3) Define the state transfer function δ: S×Ac A →S, which describes how the system goes from one state to another. Define a labeling function to associate the state s with the atomic proposition AP; (4) Define the agent's mental state set M and the mental state F extracted from the current state s a , each agent a∈A thinks F in different states a (s) = m a .

3. The method for evaluating the performance of temporal cognitive attributes of a multi-agent system according to claim 1, characterized in that: The described step (2) is specifically: (1) Define atomic formulas: First, define atomic formulas f, each of which corresponds to a basic state in the system. (2) Constructing temporal logic formulas: Use temporal logic operators to construct complex performance attribute formulas. Commonly used temporal logic operators include: "Next" (X) and "Until" (U). The specific temporal logic formula expression form is Xφ, φ1Uφ2. (3) Constructing cognitive logic formula: Introducing knowledge operators K, K a φ is used to describe the knowledge state of agent a about formula φ. Introduce knowledge operator E, E A φ represents the knowledge state that all agents in the system know φ. Introducing knowledge operators C, C A φ represents the public knowledge in the system, which means that all agents know and all agents know that other agents also know the knowledge state of φ. (4) Construct a complete performance attribute formula: Construct a performance attribute formula by combining the atomic formula, timing operator, and knowledge operator. The performance attribute formula is obtained by combining the operators addition (+), multiplication (×), conjunction (∧), disjunction (∨), minimum value (MIN), and maximum value (MAX).

4. The method for evaluating the performance of temporal cognitive attributes of a multi-agent system according to claim 1, characterized in that: The described step (3) is specifically: (1) Input the concurrent game structure model CGS of the multi-agent system, the performance attribute formula φ, and the current state s, and recursively calculate Comput(CGS, φ, s); (2) When φ is an atomic formula f, the result returns the value of the atomic formula in the current state s; (3) When φ = φ1 + φ2, the result is the sum of Comput(CGS, φ1, s) and Comput(CGS, φ2, s); (4) When φ = φ1 × φ2, the result is returned as the multiplication of Comput(CGS, φ1, s) and Comput(CGS, φ2, s); (5) When φ = φ1∧φ2, the result returns the minimum value between Comput(CGS, φ1, s) and Comput(CGS, φ2, s); (6) When φ = φ1∨φ2, the result returns the maximum value between Comput(CGS, φ1, s) and Comput(CGS, φ2, s); (7) When φ contains K a When φ is calculated, Comput(CGS,K a φ, s); (8) When φ contains E A When φ is calculated, Comput(CGS,E A φ, s); (9) When φ includes C A When φ is calculated, the call Comput(CGS,C A φ, s); (10) When φ contains MIN A When Xφ, the calculation calls Comput(CGS,MIN A Xφ, s); (11) When φ contains MAX A When Xφ, the calculation calls Comput(CGS,MAX A Xφ, s); (12) When φ contains When the calculation calls Comput (13) When φ contains When the calculation calls Comput

Images