Nonlinear Multi-objective Reinforcement Learning with Provable Guarantees

We describe RA-E3 (Reward-Aware Explicit Explore or Exploit), an algorithm with provable guarantees for solving a single or multi-objective Markov Decision Process (MDP) where we want to maximize the expected value of a nonlinear function over accumulated rewards. This allows us to model fairness-aware welfare optimization for multi-objective reinforcement learning as well as risk-aware reinforcement learning with nonlinear Von Neumann-Morgenstern utility functions in the single objective setting. RA-E3 extends the classic E3 algorithm that solves MDPs with scalar rewards and linear preferences. We first state a distinct reward-aware version of value iteration that calculates a non-stationary policy that is approximately optimal for a given model of the environment. This sub-procedure is based on an extended form of Bellman optimality for nonlinear optimization that explicitly considers time and current accumulated reward. We then describe how to use this optimization procedure in a larger algorithm that must simultaneously learn a model of the environment. The algorithm learns an approximately optimal policy in time that depends polynomially on the MDP size, desired approximation, and smoothness of the nonlinear function, and exponentially on the number of objectives.

Welfare and Fairness in Multi-objective Reinforcement Learning

We study fair multi-objective reinforcement learning in which an agent must learn a policy that simultaneously achieves high reward on multiple dimensions of a vector-valued reward. Motivated by the fair resource allocation literature, we model this as an expected welfare maximization problem, for some non-linear fair welfare function of the vector of long-term cumulative rewards. One canonical example of such a function is the Nash Social Welfare, or geometric mean, the log transform of which is also known as the Proportional Fairness objective. We show that even approximately optimal optimization of the expected Nash Social Welfare is computationally intractable even in the tabular case. Nevertheless, we provide a novel adaptation of Q-learning that combines non-linear scalarized learning updates and non-stationary action selection to learn effective policies for optimizing nonlinear welfare functions. We show that our algorithm is provably convergent, and we demonstrate experimentally that our approach outperforms techniques based on linear scalarization, mixtures of optimal linear scalarizations, or stationary action selection for the Nash Social Welfare Objective.