Explore task-driven network discovery through deep reinforcement learning in this 28-minute lecture from the Deep Learning and Combinatorial Optimization 2021 conference. Delve into the challenges of working with incomplete network data and learn how to improve network observability for more accurate analyses. Discover the Network Actor Critic (NAC) framework, which utilizes task-specific network embeddings to reduce state space complexity and learn offline policies for network discovery. Examine the performance of NAC compared to competitive online-discovery algorithms and understand the importance of planning in addressing sparse and changing reward signals. Gain insights into selective harvesting, policy functions, and modeling future rewards in network analysis. Explore real-world applications, including the control of pandemics, and get an overview of the COANET system for tackling complex network problems.
Task-Driven Network Discovery via Deep Reinforcement Learning on Embedded Spaces
Institute for Pure & Applied Mathematics (IPAM) via YouTube
Overview
Syllabus
Intro
complex networks are ubiquitous
Working with incomplete data can skew analyses
The network discovery question
A more accurate representation
Some issues that make the problem difficult
Selective harvesting via reinforcement learning
Policy function
Modeling future reward: Return function
Value function
What are current approaches missing?
State space representation
Map network states into canonical representations
Training set generation for offline learning
Episodic training
The learning objective
Our model: Network Actor Critic (NAC)
Experiments: Baselines & competitors
Experiments: Results on real data
Which graph embedding to choice?
Wrap-up: Network Actor-Critic (NAC)
Control of pandemics
Problem and high-level overview of our system: COANET
Taught by
Institute for Pure & Applied Mathematics (IPAM)
