Deep Reinforcement Learning for Fairness in Distributed Robotic Multi-type Resource Allocation

摘要

As autonomous robots are becoming a reality, we discover new challenges in coordination among these robots. We present a unique new problem that each robot makes decisions in achieving tasks that require multiple robots with different capabilities. Fair resource allocation is essential to ensure that all resource requesters acquire adequate robot resources and accomplish their tasks. We propose solutions to the fairness problem in multi-type resource allocation for multi-robot systems that have multiple resource requesters requiring heterogeneous robots with different capabilities to accomplish tasks. In particular, this work focuses on systems of single-tasking robots with multi-robot tasks (STR-MRT). In STR-MRT, the capability of a robot is the resource for accomplishing a specific task. In this problem, 1) each robot can perform only one task at a time, 2) tasks are divisible, and 3) accomplishing each task requires resources from one or more robots. We model the decentralized resource allocation in STR-MRT as a coordination game between the robots. Each robot strategically selects one resource requester. Then a consensus-based algorithm conducts formation of a robotic team for each task. We leverage the Dominant Resource Fairness (DRF) and Deep Q-learning Network (DQN) to support requester selection. The results suggest that the DQN outperforms the common Q-learning.