Deep Reinforcement Learning for Dynamic Multichannel Access

10/04/2017
1:30 pm - 4:00 pm
Expo Hall/Career Fair

Objective: Inspiration
Audience Level: All
Session Type: Poster

We consider the problem of dynamic multichannel access in a WSN. The problem can be formulated as a POMDP, which is intractable. As a solution, we implement a Deep Q-Network (DQN) that can work without any prior knowledge of the system dynamics. We show through simulations and a real testbed that DQN has the capability to learn an optimal or near-optimal policy in complex real scenarios.

Speaker(s)

Shangxing Wang, Phd Student, University of Southern California