Online Mixed-Integer Optimization in Milliseconds

We propose a method to approximate the solution of online mixed-integer optimization (MIO) problems at very high speed using machine learning. By exploiting repetitive nature optimization, we can greatly up time. Our approach encodes optimal into small amount information denoted as strategy voice framework. In this way, core part routine becomes multiclass classification problem that be solved quickly. work, extend framework real-time and high-speed applications focusing on parametric quadratic optimization. an extremely fast consisting feedforward neural network evaluation linear system where matrix has already been factorized. Therefore, does not require any solver or iterative algorithm. show proposed both in terms total computations required measured execution estimate number floating point operations completely recover function dimensions. Compared with state-of-the-art MIO routines, running time our is predictable lower than single factorization benchmark against Gurobi obtaining two three orders magnitude speedups examples from fuel cell energy management, sparse portfolio motion planning obstacle avoidance. Summary Contribution: technique learn mapping between key parameters encoding This allows us significantly improve computation resources needed solve problems. obtain simple low computing variance, which crucial settings.