A survey on discrete multi-objective reinforcement learning benchmarks

In this paper, we investigate the current state of the art in Multi- Objective Reinforcement Learning (MORL) benchmarking for problems with discrete state and action spaces. Our investigation finds that there exist a large number of different MORL benchmarks, but there seems to be little standardization among the MORL community. Through a quantative comparison, it becomes clear that many of the MORL benchmarks being used today do not have the same complexity as their Single-Objective Reinforcement Learning (SORL) counterparts. Based on the existing benchmarks for MORL problems, the authors propose an extended version of the commonly used Deep Sea Treasure (DST) benchmark. The new benchmark is shown to be flexible enough to cover a wide range of complexities, from an image-based alteration of the DST benchmark, to complexities rivalling those used in contemporary SORL research. Finally, the authors also provide an implementation of their benchmark.

Jaar van publicatie:2022

Trefwoorden:P3 Proceeding

Toegankelijkheid:Closed