Transformer Neural Network for Multi-Destination Route Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current navigation systems face challenges in efficiently determining optimized routes with multiple destinations under time constraints, particularly due to NP-hard complexity and sub-optimal solutions provided by existing deep and reinforcement learning algorithms, which struggle with real-life scenarios involving asymmetric routing distances and high numbers of visit places.
Innovation Solution
A scalable machine learning algorithm using a transformer-based encoder/decoder with embedding and reward functions is employed, focusing on route distances rather than Euclidean distances, to address the Traveling Salesman Problem with time-window constraints, enabling low response latency and adaptable performance for various scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If deep and reinforcement learning algorithms are applied to TSP, then fast solutions can be provided, but the solutions are sub-optimal and struggle with real-life scenarios
Solution Approach 1:
The patent introduces an intermediary mapping process that converts real-life road network distances into an abstract metric space where the machine learning model operates. This intermediary layer (distance matrix transformation) allows the model to learn optimal routing patterns without being constrained by the complexity of real-world asymmetric distances, thereby achieving both speed and optimality.
Solution Approach 2:
The patent transforms the problem parameters by converting actual road distances into a normalized distance matrix that preserves relative distance relationships while removing real-world constraints. This parameter transformation enables the neural network to generalize better and produce optimal solutions across different real-life scenarios while maintaining fast computation.
2Device complexity
If coordinates are used as input for route optimization, then Euclidean distance calculations are simple, but real-life routing distances are longer and asymmetric due to road networks
Solution Approach 1:
The patent performs preliminary action by pre-calculating and storing the distance matrix between all destination pairs before the actual route optimization. This pre-computed distance matrix captures real-life road network characteristics including asymmetry and longer actual distances, allowing the neural network to focus solely on learning optimal routing sequences without needing to process raw coordinate geometry during inference.
3Quantity of substance
If the number of destinations to visit is increased, then more places can be covered, but the computational complexity of finding optimal routes increases exponentially
Solution Approach 1:
The patent replaces the mechanical exhaustive search approach with a neural network-based learning system. Instead of computationally enumerating all possible route permutations (which grows factorially with the number of destinations), the system learns optimal routing patterns from training data and generalizes to new problems, achieving polynomial time complexity regardless of the number of destinations.
4Adaptability or versatility
If asymmetric routing distances are considered, then real-life one-way street limitations are addressed, but the problem becomes more complex than symmetric TSP
Solution Approach 1:
The patent designs a universal neural network architecture that handles both symmetric and asymmetric distance cases through the same learning framework. The model takes a general distance matrix as input (which can represent either symmetric or asymmetric relationships) and learns routing policies that automatically adapt to the specific characteristics of the problem instance, achieving versatility without requiring separate specialized algorithms.
Data Source
AI summary
Embodiments provide a method, a machine readable medium, a device and a vehicle for determining a route connecting a plurality of destinations in a road network, and a method, a machine readable medium and a device for training a machine learning module for determining a route connecting a plurality of destinations in a road network. The method for determining a route connecting a plurality of destinations in a road network comprises receiving input information, which comprises information on distances between the plurality of destinations in the road network. The method further comprises inputting the input information on the distances into a trained machine learning module and providing output information on the route connecting the plurality of destinations based on an output of the trained machine learning module.