CausalSim: A Causal Framework for Unbiased Trace-Driven Simulation
summary
CausalSim provides a causal model and estimator that removes bias in trace-driven simulation by learning latent network conditions from an initial randomized control trial and then counterfactually replaying traces; it maps unbiased replay to a sparse tensor completion problem and solves it with a novel completion method. Evaluation on Puffer and synthetic data builds a simulator that cuts error by 53% and 61% versus expert and supervised baselines, and a real deployment validates the counterfactual tuning.
AI-written summary · about
chapters (9)
- 0:00Introduction to trace bias
- 1:11The exogenous trace problem
- 2:12ABR as a case study
- 4:20Analyzing puffer dataset
- 7:46Modeling latent conditions
- 10:55The Causal Sim methodology
- 13:06Results and validation
- 14:08Policy optimization impact
- 16:19Conclusion and applications
related talks
19mAn Experimentation and Analytics Framework for Large-Scale AI Operations Platforms
24mTowards Smarter Intersections Through Digital Twins - The Coupling of SUMO and SCENIMINI
40mSynthetic data: when, why, and how
31mFinding the Needle in the Haystack: Predicting Storage Device Failures in...
19mCycle-Level Distribution-Based Calibration of Microscopic Intersection Models Using UAV Trajectories
17mA comparison of SUMO’s count based and countless demand generation tools
04mInterview
28mGenerate Insights, Not Excuses: How your data is slowing you down