We are investigating whether it is possible to use algorithms within machine learning to discover new optimization algorithms that function better than the current ones, which are constructed manually.
Planning, decision making and predictive control of autonomous systems including machine learning, are all based on very advanced optimization methods. The choice of optimization algorithm including its tunable parameters is still made manually even if frameworks such as disciplined convex optimization, including cvx and YALMIP, have done major progress in automatizing convex optimization.
Disciplined convex optimization is essentially a convenient way of interfacing to general purpose optimization algorithms for convex optimization. For large-scale convex optimization, distributed or not, and non-convex optimization problems the choice and tuning of the optimization algorithm is still very complicated. Also, there are several equivalent formulations of the optimization problem that impact what efficiency is obtained.
In order to overcome this major difficulty, we investigate if it is possible to directly learn these optimization algorithms and the best formulation from experiments and simulations, that is to apply and extend machine learning algorithms to derive new automated optimization algorithms that perform better than manually designed and tuned ones.
Professor Bo Wahlberg, KTH Royal Institute of Technology