Abstract: Deep learning has revolutionized scientific fields such as "automatic speech recognition" and in "image analysis" by replacing hand designed features with learnable features obtained by training a deep neural network on the raw input signal. These fields have now almost completely converted to a new paradigm where machine learning is driving their stunning progress. In this talk I will argue there is an interesting parallel in the field of OR where optimization algorithms are mostly hand designed rather than learned. The core idea of our proposal is to present many (possibly simulated) example problems and try to discover patterns in how to solve them most effectively. The machine learning tool to achieve this is called reinforcement learning where a policy is trained to maximize total future reward, where reward is defined in terms of the quality of the obtained solution. This approach has been highly successful for solving games such as GO and chess by systems such as AlphaGO and AlphaZero from Deepmind, and we claim it is also applicable to OR. By learning from many problem instances we automatically discover the features and patterns that are useful to solve new problem instances. We successfully developed such a reinforcement learning approach for a family of problems such as the TSP and the VRP which exhibit performance very close to the best hand designed systems. We thus anticipate that (partially) learnable policies for solving combinatorial optimization problems has the potential to make a significant impact in the field of OR.

Joint work with Wouter Kool.