Measuring Sample Efficiency and Generalization in Reinforcement Learning Benchmarks: NeurIPS 2020 Procgen Benchmark

• DOI: 10.48550/arxiv.2103.15332
• Publication Date: 2021-01-01
• Authors: Sharada P. Mohanty; Jyotish Poonganam; Adrien Gaidon; Andrey Kolobov; Blake Wulfe; Dipam Chakraborty; Gražvydas Šemetul...; João Schapke; Jonas Kubilius; Jurgis Pašukonis; Linas Klimas; Matthew Hausknecht; Patrick MacAlpine; Quang Nhat Tran; Thomas Tumiel; Xiaocheng Tang; Xinwei Chen; Christopher Hesse; Jacob Hilton; William H. Guss; Şahika Genç; John Schulman; Karl Cobbe
• Keywords: Benchmark (surveying); Sample (material)

The NeurIPS 2020 Procgen Competition was designed as a centralized benchmark with clearly defined tasks for measuring Sample Efficiency and Generalization in Reinforcement Learning. remains one of the most fundamental challenges deep reinforcement learning, yet we do not have enough benchmarks to measure progress community on We present design Learning which can help by doing end evaluation training rollout phases thousands user submitted code bases scalable way. top already existing Benchmark defining clear standardizing setups. aims maximize flexibility available researchers who wish future iterations such benchmarks, imposes necessary practical constraints allow system like this scale. This paper presents competition setup details analysis solutions identified through context iteration at NeurIPS.