Multi-Domain Active Learning for Multi-Agent Reinforcement Learning

Abstract

Multi-domain active learning (MDAL) is the process of simultaneously learning a number of models, each of which is specialized to execute a task in a certain domain, in order to decrease labeling effort by only utilizing the most useful data. In this paper, we propose a novel multi-domain active learning for multi-agent reinforcement learning (MARL) approach. The primary objective of the proposed approach is to enable learning with substantially less user input and provide users with the freedom to explore multiple possibilities in order to find the optimal strategy to maximize their reward. We employ the StarCraft 2 (SC2) learning environment to evaluate the effectiveness of our algorithm when comparing it with cutting-edge approaches. Our research results show that the recommended approach works better in all scenarios.