An Evolutionary Machine Learning Approach Towards Less Conservative Robust Optimization

In the recent era, multi-criteria decision making under uncertainty is gaining importance due to its wide range of applicability. Among several types of uncertainty handling techniques, Robust Optimization (RO) is considered as an efficient and tractable approach provided one has accessibility to data in uncertain regions. However, solutions of RO may actually deviate from actual results in real scenarios, due to conservative sampling. This paper proposes a methodology to amalgamate unsupervised machine learning algorithms with RO which thereby makes it data-driven. A novel evolutionary fuzzy clustering mechanism is implemented to transcript the uncertain space such that the exact regions of uncertainty are identified. Subsequently, density based boundary point detection and Delaunay triangulation based boundary construction enables intelligent Sobol based sampling in these regions for use in RO. Results of two test cases with varying dimensions are presented along with a comprehensive comparison between conventional RO approach using box uncertainty set and proposed methodology. Considered case studies include highly nonlinear real life model for continuous casting from steelmaking industries, where a time expensive multi-objective optimization problem under uncertainty is formulated to resolve the conflict in productivity and energy consumption. Optimal Artificial Neural Network (ANN) surrogate assisted optimization under uncertainty for casting model is performed to obtain solutions in realistic time. The resulting RO problem being multi-objective in nature, the Pareto solutions are obtained by NSGA II.