Configuration optimization method in multi-agent path planning

Abstract

The invention discloses a configuration optimization method in multi-agent path planning, which is used for an algorithm in path finding of agent robots such as a logistics robot and an unmanned aerial vehicle. On the basis of an original algorithm, a multi-node generation strategy is adopted to avoid falling into a local optimal solution, and the overall problem solving capacity is improved. Comprising the steps of creating an initial node Ninit, obtaining a node to be processed, comparing node configuration with target configuration, pruning invalid nodes, generating new configuration, processing the configuration, and finally outputting a found feasible solution: for a problem instance capable of being solved, returning a solution within finite time; and if the problem instance has no solution, reporting that no feasible solution exists. Aiming at the problems of poor initial solution quality, low convergence speed and the like of the existing graph path search algorithm, the method is more suitable for industrial automation application scenes with strict requirements on real-time performance and large-scale calculation.

Description

Technical Field

[0001] This application belongs to the field of industrial automation technology and is used in the path-finding algorithm of intelligent robots such as logistics robots and drones. It relates to a configuration optimization method in multi-agent path planning. Background Technology

[0002] In recent years, industrial automation technology has been widely applied and rapidly developed in fields such as warehousing and logistics, intelligent manufacturing, and autonomous driving. Its core value lies in replacing or assisting human labor through intelligent systems to achieve higher operational efficiency, lower costs, and stronger system reliability. These complex systems often require coordinating dozens, hundreds, or even thousands of autonomous mobile entities (such as robots, AGVs, and autonomous vehicles) operating simultaneously in a shared space. Therefore, how to plan efficient and collision-free movement paths for these entities has become a key technical challenge restricting system performance improvement. The Multi-Agent Path Finding (MAPF) problem is a core computational problem proposed to address this challenge.

[0003] MAPF (Multi-Active Programming) aims to plan a path from a starting point to a target point for each agent on a given map, while ensuring that all paths are collision-free in both time and space. Its optimization objective is typically to minimize the total cost of the system, such as the total number of steps or total cost of all agents. Solving the MAPF optimally is NP-hard, and its computational complexity increases dramatically with the number of agents. This makes finding high-quality or even optimal path solutions for large-scale agent swarms extremely difficult under the real-time constraints required by real-world applications.

[0004] To address this challenge, existing research primarily focuses on developing various approximation and heuristic algorithms to achieve an optimal balance between solution time and path quality. These techniques can be broadly categorized into several types: search-based methods (such as CBS), rule-based methods (such as PIBT), and methods based on compiled or mixed integer programming (such as MIP). However, existing techniques often have significant limitations: many optimal algorithms cannot scale to large-scale instances; many efficient suboptimal algorithms cannot guarantee solution quality, or even completeness; and some algorithms with asymptotic optimality often have poor initial solution quality and converge slowly to high-quality solutions.

[0005] Therefore, current technological development trends are more focused on developing MAPF algorithms that combine real-time performance, scalability, completeness, and asymptotic optimization. Summary of the Invention

[0006] This application proposes a configuration optimization method for multi-agent path planning in large-scale scenarios, building upon the framework of existing multi-agent path planning algorithms. Based on the original algorithm, a multi-node generation strategy is employed to avoid getting trapped in local optima, thereby improving the overall problem-solving capability.

[0007] The technical solution adopted in this invention is as follows:

[0008] Step 1: Input is a graph G=(V, E), where V is the graph vertex, E is the adjacency relation, and the set of agents is the agent. Initial configuration (Initial position of each agent) ), target configuration Edge cost The heuristic value is h. Here, the initial configuration S and the target configuration T represent the initial and final target position sets of all agents, respectively. Edge cost. The loss is used as a metric, specifically the total number of actions by which the agent fails to remain at the target position. Formally defined as: .

[0009] in, Defined as:

[0010] Create a new empty stack `Open` to store nodes to be processed, an empty hash table `Explored` to store processed configurations, and initialize N. gol An empty value indicates that the target node was not found.

[0011] Step 2: Create the starting node N init The configuration of the starting node, Q=S, as mentioned above, represents the initial position of all agents, with its parent node, constraints, neighbor nodes, and cost all being empty.

[0012] Step 3: Set the starting node N init Push it onto the open stack of the exploration queue, and record "Initial configuration S corresponds to node N" in Explored. init .

[0013] Step 4: Obtain the node to be processed, that is, pop an unprocessed node from the open stack, but do not pop the stack.

[0014] Step 5: Compare the configuration of node N with the target configuration T. If it matches the target configuration, record the current configuration in node N. gol It marks that a solution has been found, but does not terminate the search.

[0015] Step 6: Prune invalid nodes. Evaluate the cost of the solution generated in Step 5, comparing it to the initial heuristic value h. If the cost is greater than h, it means a solution within the cost range cannot be generated, and the search for this node is terminated. The heuristic value is the minimum cost to the target, denoted by h. h is formally defined using distance: .

[0016] Step 7: When encountering a node with empty constraints, it means all possible "successor generation rules" have been tried, and no new path can be explored; therefore, N is popped. For nodes with constraints, a constraint set is generated. The constraints refer to the situations where vertex or edge collisions occur at a certain time point. Let Q[i] represent the position of agent i∈A; a vertex collision occurs when there is a pair of agents in configuration Q that satisfy Q[i] = Q[j] (i.e., the two agents are located at the same vertex). An edge collision occurs when there is a pair of agents i≠j in two configurations X and Y that satisfy X[i] = Y[j]. When X[j]=Y[i] (i.e., the two agents exchange positions), there is an edge collision between the two configurations.

[0017] Step 8: Generate a new configuration. Using the PIBT configuration generator, input the current configuration Q and constraint C to generate Q. new .

[0018] Step 9: Process the configuration, which involves two scenarios:

[0019] If Q is not found in Explored new The record indicates a newly generated configuration, requiring the calculation of the cumulative cost of the new node. The calculation formula is as follows: That is, the cost of the new node = the cost of the current node + the edge cost between the two configurations. Then, using this node as the parent node, a new search node N is created. new It is then added to open and Explored, and finally the neighbor nodes are updated.

[0020] If there is a Q in Explored new The record is updated, the new configuration is written to Explored, the neighboring nodes are updated, and then pushed back onto the open stack.

[0021] Step 10: Output the found feasible solutions. The graph path search algorithm has completeness: for a solvable problem instance, it can return a solution in a finite amount of time; if the problem instance has no solution, it will report "no feasible solution exists".

[0022] Further, for step 4, the following two methods can also be adopted:

[0023] The first strategy is a multi-node generation strategy: When acquiring the current node to be processed, a re-search or random search is used. Instead of starting from the original node to be processed, a different node is selected. That is, when acquiring a node, there is a 1% probability of triggering a re-search or a random selection. The difference between the two methods is that the re-search removes the currently searched result and searches for other nodes; the random search means that instead of searching for the target that should be searched, a node is randomly searched in the open stack for processing.

[0024]

[0025] The second approach is a multi-path generation strategy: When searching a node, for the node with the most path collisions, the background simultaneously searches for multiple paths, setting a threshold m. Assuming the shortest path is n, if the length of the searched path is n+m, it is considered cost-acceptable, and the path of that node is updated to the newly searched path. The parameter m is:

[0026]

[0027] Further, in steps 5 and 8, there are two ways to optimize the configuration generation:

[0028] The first type:

[0029] For a graph G=(V, E), an agent i with vertex u∈V typically performs an action that involves moving to an adjacent node or remaining in place. This patent uses the distance between the moved point v and the target point p as the factor for evaluating action priority, expressed by the formula:

[0030] The design logic of the barrier is to add a barrier in the direction the higher-level agent wants to move, preventing the lower-level agent from passing through. Based on this, the distance between agent i's action v and the target point p of the higher-level agent is calculated. j The distance d between ij And the advanced intelligent agent j and its target point p j The distance d between jj Compare and use as a barrier. If d ij Greater than d jj If the action is hindered, the weight of the action will increase, and the action will be selected after comparing all surrounding higher-level agents.

[0031]

[0032]

[0033] During configuration generation, if the target position u of agent i is not equal to the current position Q[j] of agent j, and the position u is closer to the target p of agent j...j shortest path distance The distance from agent i's current position to p j shortest path distance The barrier count is incremented by 1. During the action selection phase of configuration generation, barrier sizes are compared, and actions with smaller barrier counts are prioritized.

[0034] The second type:

[0035] In the iterative configuration generation phase, an evaluation value for the agent selection loss was added, and the formal formula is as follows:

[0036]

[0037] Where v is the action (position) actually chosen by agent i, and C is the set of candidate actions for i (neighboring vertices + current vertex). Indicates the distance from position x to target i. The shortest path distance. And a new two-dimensional table R[i, v] is introduced to record the total impact on other agents when agent i chooses action v. value.

[0038] Before generating the configuration, an iterative loop is performed based on the two-dimensional table R[i, v] to gradually update the table. The number of iterations is m, which can be set to 3-5 times. During the action selection, the actions are sorted in ascending order according to the R table, and priority is given to those actions. Lower movements.

[0039] This invention addresses the problems of poor initial solution quality and slow convergence speed in existing graph path search algorithms by proposing a configuration optimization method in multi-agent path planning, which improves its performance and makes it more suitable for industrial automation applications with stringent requirements for real-time performance and large-scale computing. Attached Figure Description

[0040] Figure 1 This is a flowchart of a configuration optimization method in multi-agent path planning according to the present invention.

[0041] Figure 2 The random scenario used random-32-32-20 and random-64-64-20 maps, and executed cases with 100-450 agents (incrementing by 50 each time).

[0042] Figure 3 The complex scenario uses both warehouse-10-20-10 and warehouse-20-40-10 maps, and also executes the scenario with 100-450 intelligent agents. Detailed Implementation

[0043] The actual application environment is a factory environment, and the intelligent agents are robots, AVG vehicles, etc. This patent uses a four-connected grid map of the MAPF benchmark test set to conduct simulation experiments in two typical scenarios (random scenario and complex scenario).

[0044] The random scenario used the random-32-32-20 map and the random-64-64-20 map, and executed cases with 100-450 agents (incrementing by 50 each time). The complex scenario used the warehouse-10-20-10 map and the warehouse-20-40-10 map, and also executed cases with 100-450 agents. The complex scenario had more obstacles.

[0045] Environment Setup: This experiment requires a Linux environment. In this case, it runs on Windows 11 using WSL, specifically Ubuntu 24.04.1. Download the MAPF benchmark suite and place it in the execution folder.

[0046] The more complex the environment and the larger the scale of the intelligent agents, the more obvious the improvement effect of the method proposed in this patent.

Claims

1. A multi-agent path planning configuration optimization method, characterized by the following steps: Step 1: input is a graph G = (V, E), graph vertices V, adjacency relation E, agent set , initial configuration (initial position of each agent ), target configuration , edge cost , heuristic value h; wherein the initial configuration S and the target configuration T represent the initial position set and the final target position set of all agents respectively; the edge cost is measured in the form of loss, that is, the total number of actions of the agent not staying at the target position; the formal definition is: ; wherein is defined as: ; New empty stack Open, used to store the nodes to be processed, empty hash table Explored, used to store the processed configurations; and initialize N gol Empty, indicating that the target node is not found; Step 2: Create the start node N init including the configuration Q = S of the start node, as described earlier, with the initial position of all agents, their parent nodes, constraints, neighbor nodes, and costs being empty; Step 3: Push the start node N init into the open stack and record "start configuration S corresponds to node N init in Explored; Step 4: Obtain the node to be processed, i.e. pop an unprocessed node from the open stack, but do not pop the stack; Step 5: Compare the configuration of node N with the target configuration T, if it is the target configuration, record the configuration at this time to N gol and the flag has found a solution but does not terminate the search; Step 6: Prune invalid nodes, cost evaluation is performed on the solution generated in step 5, that is, compared with the initial heuristic value h, if greater than h, it means that the solution within the cost cannot be generated, and the search of this node is terminated; the heuristic value is the minimum cost of the current configuration to the target, denoted as h; the distance is used to formalize the definition of h: ; Step 7: If the node is empty, it means that all possible "successor generation rules" have been tried and no new path can be explored, pop N; for the node with constraints, generate the constraint set, whose constraints refer to the situation that a vertex collision or edge collision occurs at a certain time node; let Q[i] represent the position of the agent i∈A; vertex collision is that when there is a pair of agents in the configuration Q, and Q[i]=Q[j] (i.e. two agents are located at the same vertex), the configuration has vertex collision; edge collision is that when there is a pair of agents i≠j in the configurations X and Y, X[i]=Y[j] X[j]=Y[i] (i.e. two agents exchange positions), there is an edge collision between the two configurations; Step 8: Generate new configuration, using the PIBT configuration generator, input current configuration Q and constraints C, generate Q new ; Step 9: Process the configuration, which is divided into two cases: If Q is not found in Explored new The record indicates a newly generated configuration, requiring the calculation of the cumulative cost of the new node. The calculation formula is as follows: That is, the cost of the new node = the cost of the current node + the edge cost between the two configurations; then, using this node as the parent node, a new search node N is created. new And add it to open and Explored, and finally update the neighbor nodes; If there is a Q in Explored new The record is updated, the new configuration is written to Explored, the neighboring nodes are updated, and the records are pushed back onto the open stack. Step 10: Output the feasible solution found, and the graph path search algorithm has completeness: for solvable problem instances, it can return a solution within a finite time; if there is no solution to the problem instance, it will report "no feasible solution exists".

2. The multi-agent path planning configuration optimization method according to claim 1, characterized in that: Step 4 adopts a multi-node generation strategy: in obtaining the current node to be processed, a re-search or random search method is used, which does not start from the original node to be processed, but selects another node; that is, when obtaining the node, a 1% probability triggers the re-search node or randomly selects the node; the difference between the two methods is that the re-search removes the current search result and searches for other nodes; the random method means that instead of searching for the target, a node is randomly searched in the open stack for processing; 。 3. The multi-agent path planning configuration optimization method according to claim 1, characterized in that: Step 4 adopts a multi-path generation strategy: when searching for a node, for the node with the most path collisions, multiple paths are searched in the background at the same time, and a threshold m is set; assuming that the shortest path is n, if the length of the searched path is n+m, it is considered as an acceptable cost, then the path of the node is updated to the newly searched path; the parameter m is: 。 4. The multi-agent path planning configuration optimization method according to claim 1, characterized in that: In Step 5, for a graph G = (V, E), an agent i of a vertex u e V, its action usually refers to moving to an adjacent node, or staying in place; the factor used by the present application to evaluate the priority of the action is the distance between the point v after moving and the target point p, which is expressed by the formula: ; The design logic of the barrier is as follows: add a barrier in the direction that the higher-level agent wants to go to prevent the lower-level agent from passing through; based on this, the distance between agent i's action v and the target point p of the higher-level agent is determined. j The distance d between ij And the advanced intelligent agent j and its target point p j The distance d between jj For comparison, as a barrier; if d ij Greater than d jj If the action is hindered, the weight of the action will be increased, and the action will be selected after comparing all the surrounding higher-level agents. ； ； During configuration generation, if the target position u of agent i is not equal to the current position Q[j] of agent j, and the distance from position u to the target p of agent j is... j shortest path distance The distance from agent i's current position to p j shortest path distance The barrier count is incremented by 1; and during the configuration generation action selection phase, the barrier sizes are compared, and the action with the smaller barrier count is selected first.

5. The multi-agent path planning configuration optimization method according to claim 1, characterized in that: In step 5, in the iterative generation configuration phase, the evaluation value of the quantized agent selection loss is added, and the formal formula is as follows: ; Where v is the action (position) actually chosen by agent i, and C is the set of candidate actions for i (neighboring vertices + current vertex). Indicates the distance from position x to target i. The shortest path distance is determined; and a new two-dimensional table R[i, v] is introduced to record the total impact on other agents when agent i chooses action v. Values; before generating the configuration, iterate through the two-dimensional table R[i, v] to update the table step by step; the number of iterations is m, which can be set to 3-5 times; during the action selection, sort the R table in ascending order and select the action first. Lower movements.

6. The multi-agent path planning configuration optimization method according to claim 1, characterized in that: In Step 8, for a graph G = (V, E), an agent i of a vertex u e V, its action usually refers to moving to an adjacent node, or staying in place; the factor used by the present application to evaluate the priority of the action is the distance between the point v after moving and the target point p, which is expressed by the formula: ; The design logic of the barrier is to increase the barrier in the direction where the high-level agent is going to go, so as to prevent the low-level agent from passing through; based on this, the distance d between the agent i after taking the action v and the target point p of the high-level agent is compared with the distance d between the high-level agent j and the target point p of the high-level agent, as a barrier; if d is greater than d, the action hindering weight is increased, and after comparing all the high-level agents around, the action is selected again. j ij j jj ij jj ​​​​​​ ； ； During configuration generation, if the target position u of agent i is not equal to the current position Q[j] of agent j, and the distance from position u to the target p of agent j is... j shortest path distance The distance from agent i's current position to p j shortest path distance The barrier count is incremented by 1; and during the configuration generation action selection phase, the barrier sizes are compared, and the action with the smaller barrier count is selected first.

7. The multi-agent path planning configuration optimization method according to claim 1, characterized in that: At step 8, in the iterative generation configuration phase, the evaluation value of the quantized agent selection loss is added, and the formal formula is as follows: ; Where v is the action (position) actually chosen by agent i, and C is the set of candidate actions for i (neighboring vertices + current vertex). Indicates the distance from position x to target i. The shortest path distance is determined; and a new two-dimensional table R[i, v] is introduced to record the total impact on other agents when agent i chooses action v. Values; before generating the configuration, iterate through the two-dimensional table R[i, v] to update the table step by step; the number of iterations is m, which can be set to 3-5 times; during the action selection, sort the R table in ascending order and select the action first. Lower movements.

Images