NEURAL COMBINATORIAL OPTIMIZATION WITH REINFORCEMENT LEARNING
Abstract(借用pointer network用 policy gradient 优化,)
given a set of city coordinates, predicts a distribution over different city permutations.
RL(reward 总路径长度相反数) + RNN (policy gradient)
很好的解决100node 问题.
apply on KnapSack, 解决200 items
Introduction
包括两个递归神经网络（RNN）模块，编码器和解码器，两者均由长短期记忆（LSTM）单元组成
The input to the encoder network at time step i is a d-dimensional embedding of a 2D point xi,
pointing mechnism
ouputA(ref,q) 询问每一个ri被指的概率/degree
Word & Phrase
myriad applications 无数的应用
We empirically demonstrate that, 实验表明
penalize the violations of the problem’s constraints
