Sub-policy pruning in Meta Learning Shared Hierarchies

Abstract

MLSH (Meta Learning Shared Hierarchies) is a meta learning method that divide a policy into multiple sub-policies and a master-policy that picks one of the sub-policies to be actually used. By training sub-policies in advance, master policy can rapidly adjust to given environments. However, MLSH is not suitable for complicated environments. In complicated environments, a number of sub-policies are required, and it is very difficult to train them properly. We propose a method to effectively prune excessive sub-policies to give better chance the other sub-policies to betrained. We demonstrate that we can prune 40 % of sub-policies while preserving the performance.