Distance-Based ML Model for Traveling Salesman Problem Routing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing algorithms for solving the Traveling Salesman Problem (TSP) using deep and reinforcement learning rely on coordinates rather than distance, leading to sub-optimal solutions that do not accurately represent real-world routing scenarios, particularly due to asymmetric one-way routes and the complexity of road networks.
Innovation Solution
A machine-learning model is trained using reinforcement learning with a reward function based on distance to determine the optimal order of driving destinations, considering real-life road network implications, which reduces computational complexity and improves route planning efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If coordinates are used as input to train the machine-learning model, then the model can be trained using standard approaches, but the solution does not accurately represent real-world routing distances and becomes sub-optimal
Solution Approach 1:
The patent changes the input parameter from coordinates to pre-computed routing distances between location pairs. This transformation allows the model to directly learn from actual road network distances rather than geometric coordinates, improving routing accuracy while simplifying the training process by using distance matrices as input features
2Manufacturing precision
If mixed integer programming is used to solve the TSP problem, then an optimal solution can be found, but the runtime is slow and varies depending on the case
Solution Approach 1:
The patent replaces the traditional mixed integer programming mechanical optimization system with a machine-learning-based predictive system. The model is trained offline using reinforcement learning and then deployed for fast online inference, substituting the slow iterative optimization process with a fast neural network prediction that provides near-optimal solutions in real-time
3Productivity
If reinforcement learning is used to train the machine-learning model, then the model can learn optimal routing strategies, but significant effort is required for data annotation and training
Solution Approach 1:
The patent performs preliminary action by pre-computing the distance matrix between all location pairs before training the model. This preprocessing step eliminates the need for the model to learn distance calculations during training, reducing training time and computational requirements while allowing the model to focus on learning optimal routing sequences
Data Source
AI summary
Examples relate to systems, method and systems, methods and computer programs for efficiently determining an order of driving destinations and for training a machine-learning model using distance-based input data. A system for determining an order of a plurality of driving destinations is configured to obtain information on a distance between the plurality of driving destinations, the distance being defined for a plurality of routes between the plurality of driving destinations, with the routes being defined separately in both directions between each combination of driving destinations of the plurality of driving destinations within the information on the distance. The system is configured to provide the information on the distance between the plurality of driving destinations as input to a machine-learning model, the machine-learning model being trained to output information on an order of the plurality of routes based on the information on the distance provided at the input of the machine-learning model. The system is configured to determine the information on the order of the plurality of driving destinations based on the information on the order of the plurality of routes.