Comparative Synthesis: Learning Near-Optimal Network Designs by Query
summary
Framework learns an implicit objective by asking the user Compare or Propose queries and returns a near-optimal allocation after a small budget (ten) of queries. Implementation Net10Q uses random sampling plus a voting-guided learner that picks maximally informative queries, proves logarithmic (and linear under a sortable model) convergence, and calls an optimizer (Groupie) to evaluate throughput/latency. Evaluation: four real-world case studies, largest topology 114 nodes/135 links, 17-user pilot, reaches ~90% quality within six queries and ~97% after seven.
AI-written summary · about
chapters (16)
- 0:00Intro
- 0:16Network Design Problem
- 1:47What Is the Optimal Solution?
- 3:39Quantitative Synthesis Conundrum
- 4:14Programming by Example?
- 7:17Quality of Solution
- 9:43Search Space: Pareto candidate set
- 10:23Querie Informativeness: Compare
- 11:15Voting-Guided Learning Algorithm
- 11:55Convergence Guarantee
- 12:39Can the algorithm converge faster?
- 14:25Oracle-based Evaluation (Perfect Oracle)
- 15:44Pilot User Study (17 Participants)
- 16:36Pilot User Study: User Feedback
- 18:06Comparative Synthesis: Conclusion
- 23:38Optimization Scenarios
related talks
16mJust-in-Time Learning for Bottom-Up Enumerative Synthesis
17mTopoOpt: Co-optimizing Network Topology and Parallelization Strategy for Distributed...
47mRecent trends in combinatorial optimization augmented machine learning: A graph learning perspective
17mDrs.NAS: Ultra-Efficient Neural Architecture Search for Recommendation Systems
40mAdaptive query optimization in PostgreSQL
16mAccelerating Distributed Graph Learning by Using Collaborative In-Network...
39mPractical Optimization for Stats Nerds
14mHow we taught agents to use good retrieval