Towards extending algorithmic strategy planning in System Dynamics modeling

Acknowledging the importance of System Dynamics, this paper takes a first step towards extending the policy optimization capabilities, by allowing the controlled parameters to vary in time. To achieve this, for the first time a REINFORCE algorithm is being applied to a System Dynamics model. In order to improve performance, several instances are being made to compete against each other in several stages. A simple model is being used to compare this new approach to both the existing one and the optimal solution. It is shown that as expected the time varying parameters outperform the traditional constant parameter optimization and even almost reach optimality in some cases. Finally, a comparison is being drawn, when applying the same methodology to a system with noise.