Distributed Reinforcement Learning based onα-domination Strategy for Multi-criteria Decision Making and its Application to Distributed Database Systems

In the distributed systems in which information cannot be exchanged directly among agents, we deal with problems of deciding how each agent holds the shared resource. To achieve a lot of tasks greedily, agents tend to attempt to hold the resources for a long term. However the system performance decreases consequentially because it competes with the processing of other agents’ tasks. To acquire cooperative policies that avoid above competition, we formulate the resource sharing problems to multicriteria decision making problems with the priority level by using the domain knowledge into the reward. We propose distributed reinforcement learning that narrows the choice of action space by using the α-domination strategy based on value functions for the object and the cooperation. The proposed method is applied to the distributed database systems, and simulation results shows that our method acquires cooperative policies and improves the throughput performance of the system.