Stay Informed:
Baskin Engineering COVID-19 Information and Resources
Campus Roadmap to Recovery
Zoom Links: Zoom Help | Teaching with Zoom | Zoom Quick Guide

Defense: Policy Tree Based Reinforcement Learning Approaches for Medium Access

Speaker Name: 
Molly Zhang
Speaker Title: 
PhD Candidate
Speaker Organization: 
Computer Science PhD
Start Time: 
Friday, June 18, 2021 - 9:00am
End Time: 
Friday, June 18, 2021 - 10:00am
Zoom - - Passcode: 867771

Abstract: We explore reinforcement learning (RL) based approaches to create adaptive medium access control protocols that learn from past transmission history. As apposed to canonical reinforcement learning algorithms, a policy tree is used to represent both the decision space and the environment, by organizing potential transmission schedules in a binary tree. The protocol determines transmission schedule according to the policy tree, and also learns from the transmission outcome by updating the policy tree, with the goal to to maximize both channel utilization as well as fairness of the channel utilization among nodes. The updates are either editing the structure of the tree, or changing the quantitative weight of each tree node, and these two mechanisms resulted in two subset of algorithms: Adaptive Tree ALOHA and Quantitative Tree ALOHA. Both immediate and delayed acknowledgements mechanisms are created for both algorithms, which begets four families of policy tree protocols. This allows the policy tree protocols to be used in settings such as centralized wireless network as well as decentralized peer-to-peer ad-hoc networks. Policy-tree based approaches outperforms alternative MAC protocols, such as ALOHA with exponential backoff, ALOHA-Q, and deep RL methods, in terms of higher network utilization, faster learning time and high level of fairness in network bandwidth distribution.

Event Type: 
Luca de Alfaro
Graduate Program: 
Computer Science PhD