graph convolutional network for the TSP

1.解决问题：

二维的TSP问题

2.创新点：

deep Graph Convolutional Networks（build efficient TSP graph reperesentations） +a non-autoregressive manner via highly parallelized beam search

3.模型结构

3.1 Overview of the approach

1.Taking a 2D graph as input, the graph ConvNet model outputs an edge adjacency matrix denoting the probabilities of edges occurring on the TSP tour.
2. This is converted to a valid tour using beam search.

3.2 advantages of the model

1.Solution quality
2.Inference speed

3.3 model layer

3.4 emmbedding reference

4. beam search decoding

1 greedy search
2 Beam search
3 Beam search and Shortest tour heuristic

5.Hyperparameter Configurations

1.30层图卷积层
2.3层感知机
3.隐藏层为300
4.beam width b = 1280
5.k近邻的k=20

4.数据分配及分布

训练集：一百万对实例和解
验证集：10000
测试集：10000
数据仍然为分布在([0,1],[0,1])区间的数
最优解：通过求解器CONCORDE获得。

5.result：

6.generalization