Intelligent campus simulation system based on multi-agent dynamic interaction and reinforcement learning

Abstract

The invention discloses an intelligent campus simulation system based on multi-agent dynamic interaction and reinforcement learning, and belongs to the technical field of cooperation and interaction of distributed agents. Comprising a multi-agent basic simulation platform, an agent behavior module, a group influence algorithm module and a disaster scene generation module. The multi-agent basic simulation platform simulates a campus environment, provides real-time environment data for the agent behavior module through a data interface, and drives agents such as students and teachers to operate according to preset behavior logic. Meanwhile, the position of the agent and the authority level of the authority role are transmitted to the group influence algorithm module, and the group influence algorithm module calculates behavior offset in real time through a group influence algorithm, feeds back an offset instruction to the behavior module and corrects a target. Building structure parameters and various data are provided for the disaster generation module to dynamically generate a disaster model, meanwhile, a PPO algorithm is called to be combined with various statistical data to optimize a reward function, and finally an evacuation path is optimized.

Description

Technical Field

[0001] The present invention relates to an intelligent campus simulation system, belonging to the technical field of collaboration and interaction of distributed intelligent agents, and in particular to an intelligent campus simulation system based on multi-agent dynamic interaction and reinforcement learning. Background Art

[0002] Multi-Agent Systems (MAS), through the collaboration and interaction of distributed agents, have demonstrated significant potential in smart campus management. Their core value lies in achieving efficient resource scheduling and optimized emergency response through differentiated role modeling (e.g., student, teacher, security guard) and dynamic environmental perception. Existing technologies generally employ a modular design framework, breaking down the behavioral logic of different roles into independent functional modules (e.g., permission management, path planning), achieving basic collaboration through pre-set rules or centralized control.

[0003] Although the existing system divides roles into modules such as students, teachers, and security guards, the behavioral logic of the intelligent agents mostly relies on the same set of rules. Student and teacher intelligent agents adopt the same evacuation strategy in emergency scenarios, which does not reflect the teacher's organizational and management functions. In addition, the student intelligent agents lack individual personality modeling (for example, "calm" agents prioritize helping others, while "panic" agents blindly follow others), resulting in a lack of diversity in escape behavior.

[0004] Existing systems often use static disaster template libraries (e.g., preset parameter combinations for fire and earthquake). This approach makes it impossible to customize disaster chain reactions (e.g., earthquake → building collapse → electrical short circuit → secondary fire). Furthermore, environmental interactions are delayed, requiring manual reconfiguration of disaster parameters, and preventing the dynamic modification of disaster evolution logic using real-time data (e.g., sensor input). Summary of the Invention

[0005] The purpose of the present invention is to provide an intelligent campus simulation system based on multi-agent dynamic interaction and reinforcement learning. By defining the differentiated behavioral logic of students, teachers, security guards and administrators in a hierarchical manner, combined with the social psychology group influence algorithm, it simulates herd behavior and authoritative instructions, and realizes real-time linkage of multiple roles in emergency mode (such as teachers calling the roll and security guards locating people who have not been evacuated). At the same time, through multi-dimensional disaster scenario parametric modeling and reinforcement learning reward mechanism, it drives the intelligent agent to autonomously generate emergency plans, and combines the digital twin sandbox system to dynamically verify the optimization effects of parameters such as escape exit configuration and evacuation path, thereby improving the realism of daily campus management and the automated decision-making capabilities of disaster response.

[0006] The technical solution for achieving the purpose of the present invention is:

[0007] An intelligent campus simulation system based on multi-agent dynamic interaction and reinforcement learning consists of a multi-agent basic simulation platform, an agent behavior module, a group influence algorithm module, and a disaster scenario generation module. The multi-agent basic simulation platform, as the core, simulates the campus environment and provides real-time environmental data (such as agent coordinates, building layout, and time) to the agent behavior module through a data interface, driving students, teachers, and other agents to operate according to pre-set behavioral logic. Simultaneously, the agent's location and authority role authority level are transmitted to the group influence algorithm module, which calculates behavioral offsets in real time using the group influence algorithm and feeds the offset instructions back to the behavior module to correct the target. Furthermore, the disaster generation module is provided with building structural parameters (such as exit locations and volume) and various data (such as temperature sensor data and smoke concentration data) to dynamically generate disaster models. The PPO algorithm, combined with statistically analyzed data, optimizes the reward function to ultimately optimize the evacuation path.

[0008] Furthermore, the multi-agent basic simulation platform is based on the open source project AI-Town, which is used to build virtual campus scenes and manage the basic behavioral logic of student, teacher, security and administrator agents (such as the memory function of the agent), and is combined with other modules to realize the simulation of various agents, campus scenes and disaster scenes.

[0009] Furthermore, the agent behavior module includes student agents, teacher agents, security agents and administrator agents. For student agents, the personality, daily behavior logic and learning habits (such as whether they skip classes, whether they study on their own, etc.) are defined; for teacher agents, the teaching style (such as: relaxed and humorous, strict, etc.) is defined and the classroom management strategy (such as: the frequency of roll call in class, the amount of homework, etc.) is adjusted according to the teaching style; for security agents, the daily patrol sections are defined and the optimal escape route is dynamically planned through the A* algorithm when a disaster occurs; for administrator agents, their daily tasks are defined (such as: supervising whether a disaster occurs, etc.) and disaster evacuation warnings are activated in a timely manner.

[0010] Furthermore, the group influence algorithm module calculates the individual behavior offset through the explicit authority instruction weight (teacher instruction priority) and implicit conformity effect (crowd density attraction) in social psychology. The mathematical expression is:

[0011]

[0012] The first item is the explicit authority influence item, α is the global weight coefficient of authority influence (specifically set according to the scenario), A is the instruction intensity of the authority role (for example, the higher the teacher's authority, the greater the value), D cis the real-time physical distance between the individual and the authority role, the second term is the implicit conformity effect term, β is the global weight coefficient of the conformity effect (lower for calmer individuals), N is the crowd density within the individual's field of view, D m It is the maximum effective distance of an individual's field of vision (such as the length of a classroom).

[0013] Furthermore, the disaster scenario generation module constructs highly configurable disaster scenarios (such as fire, earthquake, etc.), and drives the agent to generate and optimize emergency plans through mathematical modeling and reinforcement learning algorithms, and trains the agent to generate evacuation paths through the PPO algorithm. The reward function is expressed as:

[0014] R=ω1·N-ω2·T-ω3·C

[0015] Where N is the number of people successfully evacuated, T is the average evacuation time, C is the number of collisions (the number of times the agent collides with other agents or obstacles during the evacuation process), and ω1, ω2, and ω3 are the weights of each factor.

[0016] Compared with the prior art, the present invention has the following technical effects:

[0017] Differentiated role behavior modeling and dynamic response mechanisms address the behavioral homogeneity of existing agents by defining differentiated behavior logic for students, teachers, security guards, and administrators. For example, student agents generate different evacuation plans and routes based on personality parameters (e.g., a "calm" agent follows the teacher's instructions and evacuates safely, while a "panic" agent panics and triggers herd behavior). Teacher agents also adjust classroom management strategies based on their teaching style (e.g., relaxed or strict), significantly enhancing the simulation's realism and decision-making diversity. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a hierarchical diagram of the overall system.

[0019] Figure 2 This is a simulation of the campus. DETAILED DESCRIPTION

[0020] The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0021] In-depth application of social psychology. The group influence algorithm module achieves precise intervention on individual behavior through the dynamic balance between explicit authority instructions and implicit conformity effect. Formula In the model, the authority role instruction intensity A is associated with the authority level, and the conformity coefficient β is adaptively adjusted according to the individual personality, so that the group's behavior can both respond to the authority's instructions and conform to the conformity effect in real society.

[0022] Dynamically generate evacuation routes based on disaster scenarios. The agent is trained using the PPO algorithm and generates the optimal evacuation path by maximizing the reward function based on the formula R = ω1·N-ω2·T-ω3·C, taking into account the number of successfully evacuated people, the average evacuation time, and the number of collisions.

[0023] In order to better understand the present invention, the following examples are provided for implementation of the present invention.

[0024] Fire scenario: The administrator agent monitors the temperature sensor data of a certain teaching building in real time and finds that the data exceeds 80℃ and the smoke concentration exceeds 500ppm. It immediately triggers the fire scenario generation module and sends evacuation instructions to the teacher and student agents through the broadcast system. At this time, the teacher agent triggers the emergency roll call function according to the preset authority level A=0.9 and starts to lead the students to evacuate. The group influence algorithm starts to calculate the student behavior offset: the calm student (β=0.3) is 5 meters away from the teacher (D c =5), authority influence item (D m =15m), driving them to follow the teacher toward the safe exit. Meanwhile, the panicked student (β = 0.7), triggered by the crowd density N = 30, follows the crowd toward the nearest exit. Meanwhile, the security agent uses the A* algorithm to generate the shortest global escape route and adjusts it in real time. The PPO algorithm optimizes the strategy based on the reward function R = 2.6N - 0.01TC, guiding the agent to avoid obstacles.

[0025] Daily management: At 8 a.m., after a strict teacher (strictness index of 0.9) enters the classroom, he or she triggers a roll call for each class (50 minutes apart) and records the roll call. When assigning homework, the system assigns homework to the class by retrieving the skipping records of all student agents in the class this week: if the class student agent attendance rate is high, the homework load is small, homework load = 3 + 2 (average number of absences 0.5). At the same time, the extroverted student's memory retrieves "last week's study efficiency was less than 50%", triggering the model to adjust today's study time from 2 hours to 1 hour, and switch the route from the study room to the playground for socializing.

Claims

1. An intelligent campus simulation system based on multi-agent dynamic interaction and reinforcement learning, characterized by: The system includes a multi-agent basic simulation platform, an agent behavior module, a group influence algorithm module, and a disaster scenario generation module. The multi-agent basic simulation platform serves as the core, simulating the campus environment and providing real-time environmental data to the agent behavior module through a data interface, driving student and teacher agents to operate according to preset behavioral logic. At the same time, the agent's location and authority role permission level are transmitted to the group influence algorithm module, which calculates the behavior offset in real time through the group influence algorithm, feeds the offset instruction back to the behavior module, and corrects the target. The disaster scenario generation module is provided with building structure parameters and various data to dynamically generate a disaster model. At the same time, the PPO algorithm is called to optimize the reward function in combination with various statistical data, and ultimately the evacuation path is optimized.

2. The intelligent campus simulation system based on multi-agent dynamic interaction and reinforcement learning according to claim 1 is characterized in that: The multi-agent basic simulation platform is based on the open source project AI-Town, which is used to build virtual campus scenes and manage the basic behavioral logic of student, teacher, security and administrator agents, and is combined with other modules to realize the simulation of various agents, campus scenes and disaster scenes.

3. The intelligent campus simulation system based on multi-agent dynamic interaction and reinforcement learning according to claim 1 is characterized in that: The agent behavior module includes student agents, teacher agents, security agents and administrator agents. For student agents, personality, daily behavior logic and learning habits are defined; for teacher agents, teaching styles are defined and classroom management strategies are adjusted according to the teaching styles; for security agents, daily patrol sections are defined and the optimal escape route is dynamically planned using the A* algorithm when a disaster occurs; for administrator agents, their daily tasks are defined and disaster evacuation warnings are activated in a timely manner.

4. The intelligent campus simulation system based on multi-agent dynamic interaction and reinforcement learning according to claim 1 is characterized in that: The group influence algorithm module calculates the individual behavior offset through the explicit authority instruction weight and implicit conformity effect in social psychology. The mathematical expression is: The first item is the explicit authority influence item, α is the global weight coefficient of authority influence, A is the instruction strength of the authority role, D c is the real-time physical distance between the individual and the authority role; the second term is the implicit conformity effect term, β is the global weight coefficient of the conformity effect, N is the crowd density within the individual's field of view, D m The maximum effective distance of an individual's field of vision.

5. The intelligent campus simulation system based on multi-agent dynamic interaction and reinforcement learning according to claim 1 is characterized in that: The disaster scenario generation module constructs highly configurable disaster scenarios and drives the agent to generate and optimize emergency plans through mathematical modeling and reinforcement learning algorithms. It also trains the agent to generate evacuation paths through the PPO algorithm. The reward function is expressed as: R=ω1·N-ω2·T-ω3·C Where N is the number of people successfully evacuated, T is the average evacuation time, C is the number of collisions, and ω1, ω2, and ω3 are the weights of each factor.

6. The intelligent campus simulation system based on multi-agent dynamic interaction and reinforcement learning according to claim 1 is characterized in that: The group influence algorithm module achieves precise intervention on individual behavior through the dynamic balance between explicit authority instructions and implicit conformity effect; formula In the model, the authority role instruction intensity A is associated with the authority level, and the conformity coefficient β is adaptively adjusted according to the individual personality, so that the group's behavior can both respond to the authority's instructions and conform to the conformity effect in real society.

7. The intelligent campus simulation system based on multi-agent dynamic interaction and reinforcement learning according to claim 1 is characterized in that: Dynamically generate evacuation routes based on disaster scenarios; train the intelligent agent through the PPO algorithm and use the formula R=ω1·N-ω2·T-ω3·C to comprehensively consider the successful evacuation of people, average evacuation time and number of collisions and maximize the reward function to generate the optimal evacuation path.

Images