Tackling Decision Processes with Non-Cumulative Objectives using Reinforcement Learning
This work introduces a method to solve Non-cumulative Markov Decision Processes (NCMDPs), where the goal is maximizing an arbitrary function of rewards, not just their sum. It creates a mapping that transforms these complex problems into standard MDPs, allowing existing reinforcement learning solvers to be used directly. This expands RL's applicability to risk-adjusted metrics and improves performance in areas like portfolio optimization.