Online stochastic optimization under time constraints

This paper considers online stochastic optimization problems where uncertainties are characterized by a distribution that can be sampled and where time constraints severely limit the number of offline optimizations which can be performed at decision time and/or in between decisions. It proposes a generic framework for online stochastic optimization and several of its instantiations. In particular, it studies the expectation algorithm E that evaluates all choices against all samples at each decision step and introduces the consensus C and regret R algorithms that only solve each sample once per step. Both theoretical and experimental results are presented on the algorithms. The theoretical results indicate that, under reasonable and practical assumptions, the expected quality loss of algorithm E is o(1), while algorithm R provides a (1 + o(1))approximation when its underlying regret function is a -approximation. The experimental results are presented on three problems of fundamentally d...

• Type: Journal
• Year: 2010
• Where: ANOR
• Authors: Pascal Van Hentenryck, Russell Bent, Eli Upfal
• Keywords: Algorithms, ANOR 2007, ANOR 2010, Online Stochastic Optimization, Stochastic Optimization Problems