"Active Learning for Convex Projection"Kovacova, GabrielaConvex projection problems appear within the context of computing reachable sets, solving multi-objective control problems via set-valued Bellman's principle, computing Nash equilibria or computing systemic risk. The state-of-the-art literature contains inner and outer approximation algorithms for solving convex projection based on geometrical considerations and iterative updating using supporting hyperplanes. A connection between convex projection and convex multi-objective optimization is also known. \\ Recently, neural networks have been successfully used to solve convex multi-objective problems. This inspired us to explore a machine learning approach to convex projection problems as well. In this talk we present an active learning approach to learning inner an outer approximation of the convex projection target set through a pair of neural networks. Joint work with Zachary Feinstein (Stevens Institute of technology) |
« back