Graph Convolutional Neural Network Path Planning for Mandatory Nodes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining optimal paths in graphs with a large number of nodes and mandatory nodes are inefficient, leading to long resolution times or sub-optimum solutions, making it difficult to find the best path for autonomous vehicles in complex environments.
Innovation Solution
A method utilizing a graph convolutional neural network (GCNN) to obtain an initial order of travel through mandatory nodes, followed by local optimization to define a path that minimizes total transition costs, allowing for almost optimal path determination in reasonable time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If exhaustive testing of all possible node combinations is used to find the optimal path, then the solution quality is optimal, but the resolution time becomes unacceptably long for graphs with large numbers of nodes and mandatory nodes
Solution Approach 1:
The patent applies preliminary action by using a Graph Convolutional Neural Network (GCNN) to pre-compute and predict an initial optimal order of mandatory nodes before performing any optimization. The GCNN is trained offline to learn the optimal ordering strategy, so when a new pathfinding instance is presented, the system already has a high-quality initial solution that requires minimal refinement, thus achieving optimal solution quality with very fast resolution time.
Solution Approach 2:
The patent replaces the mechanical exhaustive search system with a neural network-based predictive system. Instead of systematically testing all possible node combinations (mechanical enumeration), the GCNN directly predicts the optimal node ordering based on learned patterns from training data. This substitution transforms the computational approach from brute-force mechanical search to intelligent predictive reasoning, achieving both speed and accuracy.
2Productivity
If known optimization methods such as 2-opt, 3-opt, or Lin-Kernighan heuristic are used to reduce resolution time, then the resolution time decreases, but the solution quality becomes highly sub-optimum
Solution Approach 1:
The GCNN performs preliminary action by predicting an initial node ordering that is already close to optimal, eliminating the need for extensive optimization iterations. Traditional methods start from random or heuristic orderings and require many optimization cycles to improve quality, whereas the GCNN provides a high-quality starting point that requires minimal refinement, thus achieving both fast resolution and optimal solution quality simultaneously.
Solution Approach 2:
The patent changes the fundamental parameter of how the initial solution is generated. Instead of using random initialization or simple heuristics followed by iterative optimization, the system uses a trained neural network to directly generate a high-quality initial ordering. This parameter change in the initialization strategy transforms the entire optimization trajectory, allowing the system to reach optimal solutions much faster with far fewer iterations.
3Adaptability or versatility
If the graph contains a large number of mandatory nodes that must be visited, then the path complexity increases, but the resolution time becomes prohibitively long using traditional methods
Solution Approach 1:
The GCNN is specifically trained to handle instances with multiple mandatory nodes by learning the optimal ordering strategy during offline training. When presented with a new instance containing many mandatory nodes, the pre-trained network immediately predicts the correct ordering without being overwhelmed by the complexity, thus achieving fast resolution times even for instances with large numbers of mandatory nodes that would be intractable for traditional methods.
Solution Approach 2:
The patent uses the GCNN to create a copied or predicted version of the optimal node ordering based on patterns learned from training instances. Instead of solving the complex multi-mandatory-node problem from scratch, the system copies the successful ordering strategy learned during training and adapts it to the new instance, achieving fast and accurate solutions for complex instances with many mandatory nodes.
Data Source
AI summary
A path is defined to be followed by a vehicle in an environment represented by a graph of nodes connected by edges. Each node represents the vehicle's position. Each edge between two nodes indicates a transition cost between the two nodes. An instance of the graph, which is represented by a start node, a destination node and mandatory nodes through which the vehicle must pass, is obtained. An initial order of travel through the mandatory nodes is obtained by applying a procedure, using a graph convolutional neural network, adapted to the graph, to the instance. A procedure of local optimisation of the mandatory nodes is executed in order to obtain an optimised order. A path for the instance is defined to be resolved: from the optimised order and for each pair of nodes of the instance to be resolved, from a shortest path between the nodes of the pair.