Abstract for:A Competitive Analysis of a Smart Optimization Framework
This paper introduces and analyzes a new optimization framework based on deep Reinforcement Learning for System Dynamics models as interfaces to complex dynamic systems. It extends the traditional optimization capabilities on these models significantly by allowing parameters to vary in time, while practicability is kept in mind. To this end, the first application of a REINFORCE algorithm with neural networks to such models is being designed and analyzed. To increase performance, a new meta-algorithm for variance reduction is then introduced. Furthermore, it is demonstrated how methods from online algorithm analysis can be used to analyze such frameworks for the optimization of System Dynamics models in detail.