Abstract
Identifying optimization in a system from any discipline requires identifying ways to either maximize or minimize objectives of interest-or even trade-offs between these choices where goals must be balanced. Traditionally, optimal control theory has been used specifically for applications that are represented by ordinary differential equations. Here, we introduce a new approach to optimization that can be applied not only to ordinary differential-equation based systems, but importantly to other more complex models, such as agent-based model systems. To this end, we create a novel machine learning optimization pipeline that uses a Kriging-based surrogate model to predict objective functions. We use a Pareto optimization algorithm to identify regimens that maximize improvement to the predicted optimal set and then rank these findings. As an example, we apply this to a model system that captures drug treatment of hosts during infection with Mycobacterium tuberculosis. Typically for treatment of tuberculosis, a multiple drug regimen is used where four antibiotics are administered for a lengthy time frame of 6-9 months. We apply our new method to optimize treatment in the face of many choices for drugs, combinations and dosages and link for the first time with rankings to the optimized set of outcomes.