Outcome Weighted Learning in Dynamic Treatment Regimes

Abstract

This chapter discusses applications of information geometry in a paradigm of reinforcement learning with emphasis on dynamic treatment regimes which have progressed recently in the learning algorithm with outcome weighed learning. The probabilistic framework for a triple expressing state, action, and reward is formulated in multiple stages, in which a decision function defined by a state and action is estimated to make an optimal policy for a given dataset. Decision consistency for a decision function is introduced by the state-value function in the space of all the decision functions. We introduce the $\varPsi$-divergence on the decision function space with a generator function $\varPsi$, and investigate statistical properties for the $\varPsi$-loss function conducted by $\varPsi$-divergence. An outcome-weighed learning algorithm for the decision function is considered in a boosting approach in association with the prediction in supervised learning.