Multi-Objective Deep Inverse Reinforcement Learning through Direct Weights and Rewards Estimation

Inverse reinforcement learning (IRL) methods effectively imitate complex behaviors and analyze intentions by estimating rewards from expert demonstrations. However, while most real-world problems are multi-objective problems (MOPs), most IRL methods target the estimation of a single-objective reward, thus creating a need for multi-objective IRL (MOIRL) methods to estimate rewards for MOPs. Previous MOIRL methods used matrix factorization, an indirect estimation method with a two-step learning process. We propose a novel MODIRL framework that directly estimates weights and rewards in MOIRL. Further, we experimentally verify that the proposed method can simultaneously estimate both: appropriate weights and rewards.