Model-Free Interpretable Trust Region Policy Optimization Algorithm

To make machine learning widely applicable in today's society, we need to create models that are understood by laypeople. Creation of techniques that provide decision support would be highly beneficial in areas such as healthcare, finance, and robotics. Towards that goal, we design a model-free, interpretable version of the Trust Region Policy Optimization method. The method works by creating policies that have the structure of a differentiable decision tree. We lay out the construction of the method and list the subproblems that have to be solved at each leaf and internal node. We also provide convergence bounds to get this algorithm to work effectively.