MineRL NeurIPS 2020 Competition | Technology Org
The MineRL 2020 Competitiveness aims to foster the growth of algorithms which can successfully leverage human demonstrations to significantly lower the range of samples desired to fix complicated, hierarchical, and sparse environments.
To that conclusion, participants will contend to create units that can receive a diamond in Minecraft from uncooked pixels utilizing only 8,000,000 samples from the MineRL simulator and four days of training on a solitary GPU device. Individuals will be provided the MineRL-v0 dataset (site, paper), a significant-scale assortment of over 60 million frames of human demonstrations, enabling them to make the most of professional trajectories to limit their algorithm’s interactions with the Minecraft simulator. More detailed track record on the levels of competition and its structure can be observed in the MineRL 2020: NeurIPS Competitiveness Proposal.
The job of the levels of competition is resolving the MineRLObtainDiamondVectorObf-v0 environment. In this setting, the agent begins in a random starting off area without any products, and is tasked with obtaining a diamond. This job can only be completed by navigating the complicated merchandise hierarchy of Minecraft.
The agent gets a large reward for obtaining a diamond as very well as smaller sized, auxiliary benefits for obtaining prerequisite products. In addition to the main setting, we offer a range of auxiliary environments. These consist of duties which are either subtasks of ObtainDiamond or other duties inside Minecraft.