Arena: A Scalable and Configurable Benchmark for Policy Learning

Sirui Xu, Shuang Liu, Tongzhou Mu, Zhiwei Jia, Yiran Wu, Hao Su

Keywords: GNN-based policy, object-based state representation, scalable policy, configurable benchmark

TL;DR: We introduce the Arena benchmark, a scalable and configurable benchmark for policy learning.

Abstract: We believe current benchmarks for policy learning lack two important properties: scalability and configurability. The growing literature on modeling policies as graph neural networks calls for an object-based benchmark where the number of objects can be arbitrarily scaled and the mechanics can be freely configured. We introduce the Arena benchmark, a scalable and configurable benchmark for policy learning. Arena provides an object-based game-like environment where the number of objects can be arbitrarily scaled and the mechanics can be configured with a large degree of freedom. In this way, arena is designed to be an all-in-one environment that uses scaling and configuration to smoothly interpolates multiple dimensions of decision making that require different degrees of inductive bias.

URL: https://github.com/Sirui-Xu/Arena