Date of Award

May 2018

Degree Type


Degree Name

Doctor of Philosophy (PhD)


Electrical Engineering and Computer Science


Jae C. Oh


Deep Q Learning, Fair Resource Allocation, Multi-robot System, Reinforcement Learning

Subject Categories



Fair resource allocation is essential to ensure that all resource requesters acquire adequate resources and accomplish tasks. We propose solutions to the fairness problem in multi-type resource allocation for multi-robot systems that have multiple resource requesters. We apply the dominant resource fairness (DRF) principle in our solutions to two different systems: single-tasking robots with multi-robot tasks (STR-MRT) and multi-tasking robots with single-robot tasks (MTR-SRT). In STR-MRT, each robot can perform only one task at a time, tasks are divisible, and accomplishing each task requires one or more robots. In MTR-SRT, each robot can perform multiple tasks at a time, tasks are not divisible, and accomplishing each task requires only one robot.

We present centralized solutions to the fairness problem in STR-MRT. Meanwhile, we model the decentralized resource allocation in STR-MRT as a coordination game between the robots. Each robot subgroup is formed by robots that strategically select the same resource requester. For a requester associated with a specific subgroup, a consensus-based team formation algorithm further chooses the minimal set of robots to accomplish the task. We leverage the Deep Q-learning Network (DQN) to support requester selection. The results suggest that the DQN outperforms the commonly used Q-learning.

Finally, we propose two decentralized solutions to promote fair resource allocation in MTR-SRT, as a centralized solution already exists. We first propose a task-forwarding solution in which the robots need to negotiate the placement of each task. In our second solution, each robot first selects resource requesters and then independently allocates resources to tasks that arrive from the selected requesters. The resource-requester selection phase of the latter solution models a coordination game that is solved by reinforcement learning. The experimental results suggest that both approaches outperform their baselines.


Open Access

Included in

Engineering Commons