Ahmed Stohy,Heba-Tullah Abdelhakam,Sayed Ali,Mohammed Elhenawy,Abdallah A Hassan,Mahmoud Masoud,Sebastien Glaser,Andry Rakotonirainy
Abstract:In this work, a novel idea is presented for combinatorial optimization problems, a hybrid network, which results in a superior outcome. We applied this method to graph pointer networks [1], expanding its capabilities to a higher level. We proposed a hybrid pointer network (HPN) to solve the travelling salesman problem trained by reinforcement learning. Furthermore, HPN builds upon graph pointer networks which is an extension of pointer networks with an additional graph embedding layer. HPN outperforms the graph pointer network in solution quality due to the hybrid encoder, which provides our model with a verity encoding type, allowing our model to converge to a better policy. Our network significantly outperforms the original graph pointer network for small and large-scale problems increasing its performance for TSP50 from 5.959 to 5.706 without utilizing 2opt, Pointer networks, Attention model, and a wide range of models, producing results comparable to highly tuned and specialized algorithms. We make our data, models, and code publicly available [2].
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the classic Travelling Salesman Problem (TSP), which is a combinatorial optimization problem. Specifically, the goal of TSP is, given a set of cities and the distances between each pair of them, to find the shortest path such that a travelling salesman starts from one city, visits each city exactly once and then returns to the starting point, with the total travel distance being the shortest.
### Main Problems and Challenges
1. **NP - hard Problem**: TSP is an NP - hard problem, which means that it is very difficult to find the optimal solution in polynomial time.
2. **Limitations of Existing Methods**: Although there are many approximation algorithms and heuristic algorithms (such as Christofides' algorithm, local search and Lin - Kernighan heuristic algorithm) that can handle TSP, these methods still have problems such as high computational cost and insufficient accuracy when facing large - scale problems.
3. **Generalization Ability of Deep Learning Models**: Existing neural combinatorial optimization models (such as Pointer Networks, Attention Models, etc.) have shown certain potential, but when dealing with large - scale problems, their generalization ability and accuracy still need to be improved.
### Solutions Proposed in the Paper
To solve the above problems, the author proposes a new Hybrid Pointer Network (HPN). The main innovations of HPN include:
- **Hybrid Encoder**: Combines the Transformer encoder and the graph embedding layer to extract a richer feature representation.
- **Multi - decoder Structure**: Introduces an additional decoding layer, enhancing the decision - making ability of the model.
- **Reinforcement Learning Training**: Uses the Actor - Critic algorithm for training, improving the convergence speed and performance of the model.
Through these improvements, HPN has achieved results significantly better than existing methods on both small - scale and large - scale TSP problems. In particular, without using post - processing techniques such as 2 - opt, HPN can still obtain better solutions.
### Summary
This paper aims to solve the optimization problems in TSP by proposing a new Hybrid Pointer Network (HPN), especially in view of the shortcomings of existing methods in dealing with large - scale problems. The experimental results show that HPN not only outperforms existing deep - learning models in performance, but also performs well in generalization ability and practical applications.