Simulated annealing for optimization of graphs and sequences

Optimization of discrete structures aims at generating a new structure with the better property given an existing one, which is fundamental problem in machine learning. Different from continuous optimization, realistic applications optimization (e.g., text generation) are very challenging due to complex and long-range constraints, including both syntax semantics, structures. In this work, we present SAGS, novel Simulated Annealing framework for Graph Sequence optimization. The key idea integrate powerful neural networks into metaheuristics simulated annealing, SA) restrict search space We start by defining sophisticated objective function, involving interest pre-defined constraints grammar validity). SAGS searches towards performing sequence local edits, where deep generative propose editing content thus can control quality editing. evaluate on paraphrase generation molecule graph respectively. Extensive results show that our approach achieves state-of-the-art performance compared methods terms automatic human evaluations. Further, also significantly outperforms all previous generation.