Trustworthy decision making with sustainability

Abstract

Artificial intelligence (AI) has been advancing rapidly, demonstrating remarkable success in helping decision making in complicated scenarios. However, the black-box nature of machine learning (ML) models raises the important question of trustworthy decisions, while the significant cost in adopting ML-based new strategies could potentially influence the sustainability. In this thesis, we study problems related to sustainability and trustworthiness in decision making with AI. We begin with sustainability. In offline reinforcement learning (RL), there could be a significant policy switching cost whenever the concerned agent wants to adopt a new policy, which potentially makes the policy switching less sustainable. We initiate the systematic study of such problem setting with a mathematical framework and propose an actor-critic algorithm to find polices that better balance between the return and the cost. Then we extend such setting to online RL with a theoretical framework for cost-aware policy learning. We also tackle the state variable selection problem in RL, where we introduce the new definition of minimal sufficient state as the subset of state variables that affect the return, and propose a sequential variable selection method, SEEK, to efficiently select state variables. Such selection not only could help the interpretability of policies, making the decision more trustworthy, but also could reduce the training cost in downstream policy learning.