Multi-objective Reinforcement Learning with Path Integral Policy Improvement

Multi-objective reinforcement learning (MORL) for robot motion learning is a challenging problem not only because of the scarcity of the data but also of the high-dimensional and continuous state and action spaces. Most existing MORL algorithms are inadequate in this regard. However, in single-objective reinforcement learning, policy-based algorithms have solved the problem of high-dimensional and continuous state and action spaces. Among such algorithms is policy improvement with path integral (PI2), which has been successful in robot motion learning. P$\mathrm{I}^{2}$ is similar to evolution strategies (ES), and multi-objective optimization is a hot topic in ES algorithms. This paper proposes a MORL algorithm based on P$\mathrm{I}^{2}$ and multi-objective ES, which can handle the problem related to robot motion learning. The effectiveness is shown via numerical simulations.