The author proposes CoMOGA, a novel CMORL algorithm that transforms objectives into constraints, ensuring invariance to objective scales and stable constraint handling.
The authors propose a multi-objective actor-critic (MOAC) algorithmic framework that provably converges to a Pareto-stationary solution in finite time and with bounded sample complexity, for both discounted and average reward settings in multi-objective reinforcement learning (MORL) problems.
This research paper introduces LC-MOPG, a novel algorithm for Multi-Objective Reinforcement Learning (MORL) that utilizes a latent-conditioned policy gradient approach to efficiently approximate the Pareto frontier and discover diverse Pareto-optimal policies by training a single neural network.