Strategy Generation Under Uncertainty Using Bayesian Networks and Black Box Optimization

We describe a mechanism for optimal strategy generation from a Bayesian belief network (BBN). This system takes a BBN model either created by the user or derived from data. The user then specifies a set of goals (consisting of both objectives and constraints) and the observed and actionable variables in the model. The system then applies an optimizer to develop strategies that optimally achieve the specified goals. The system can be used by either human decision makers or autonomous agents. A distinguishing feature of the system is the ability to return strategies in the form of deterministic actions that result in the highest probability of achieving the desired goals. This allows the user to execute the strategies without further reasoning. In this paper we describe the architecture of the system and show examples of developing strategies from models created either by domain experts or directly from data