- Abstract: This paper introduces a framework for solving combinatorial optimization problems by learning from input-output examples of optimization problems. We introduce a new memory augmented neural model in which the memory is not resettable (i.e the information stored in the memory after processing an input example is kept for the next seen examples). We used deep reinforcement learning to train a memory controller agent to store useful memories. Our model was able to outperform hand-crafted solver on Binary Linear Programming (Binary LP). The proposed model is tested on different Binary LP instances with large number of variables (up to 1000 variables) and constrains (up to 700 constrains).
- TL;DR: We propose a memory network model to solve Binary LP instances where the memory information is perseved for long-term use.
- Keywords: Memory Networks, Combinatorial Optimization, Binary LP