The Andrew and Erna Viterbi Faculty of Electrical & Computer Engineering
Combinatorial Optimization with Graph Reinforcement Learning
Combinatorial optimization problems are abundant in machine learning applications. In this talk we show how graph neural networks can exploit the structure of combinatorial optimization problems to learn their solutions. We begin by showing how graph neural networks with evolution strategy search in the neural network computational graphs can be used in deep neural network accelerators for automated memory mapping instead of manual heuristic approaches. We train and validate our approach directly on the Intel NNP-I chip for inference and show that our method outperforms policy-gradient, evolutionary search and dynamic programming baselines on BERT, ResNet-101 and ResNet-50. We additionally achieve 28-78% speed-up compared to the native NNP-I compiler on all three workloads. If time permits, we will also explore how to use reinforcement learning and graph neural networks to solve generic combinatorial problems by learning to select cutting-planes in SCIP (Solving Constraint Integer Programs) to minimize its primal-dual gap in shorter time.
* M.Sc. Under the supervision of Prof. Shie Mannor and Tamir Hazan.
Zoom link: https://technion.zoom.us/j/97640654068
Thu 22 Jul 2021
Start Time: 10:00
End Time: 11:00
Zoom meeting | The Andrew And Erna Viterbi Faculty Of Electrical & Computer Engineering