Learning and anticipation in online dynamic optimization with evolutionary algorithms
Peter A. N. Bosman, Han La Poutré · 2007
The focus of this paper is on how to design evolutionaryalgorithms (EAs) for solving stochastic dynamicoptimization problems online, i.e.~as time goes by.For a proper design, the EA must not only be capableof tracking shifting optima, it must also take intoaccount the future consequences of the evolveddecisions or actions. A previousframework describes how to build such EAs in thecase of non-stochastic problems. Most real-worldproblems however are stochastic. In this paper weshow how this framework can be extended to properlytackle stochasticity. We point out how thisnaturally leads to evolving strategiesrather than explicit decisions. We formalizeour approach in a new framework. The newframework and the various sourcesof problem-difficulty at hand are illustratedwith a running example. We also apply ourframework to inventory management problems, an importantreal-world application area in logistics. Our results show,as a proof of principle, the feasibility and benefitsof our novel approach.