Predictive Response Optimization: Using Reinforcement Learning to...
summary
A system called Predictive Response Optimization combines reinforcement learning with model predictive control to pick enforcement actions per account using abuse and cost models, then tunes global tradeoff weights by simulation. It deploys on Instagram and Facebook and reduces weighted scraping requests by 59% and 4.5% respectively with no statistically significant user-impact, adapts automatically to new actions, bugs, and adversarial drift, and shows case studies like forced reauth cutting requests 3%, days-active +68%, SMS spend −80%.
AI-written summary · about
chapters (9)
- 0:00Predictive response optimization intro
- 1:14Traditional enforcement limits
- 3:01Action selection perspective
- 4:47How PRO works
- 7:14Model predictive control in PRO
- 8:15Real-world results
- 9:39Adapting business objectives
- 10:22New actions and systemic changes
- 11:25Conclusion and future
related talks
27mUsing Reinforcement Learning to Optimize IAP Offer Recommendations in Mobile Games
40mNext Best Action: Maximizing Customer Value
12mMachine Learning for Adaptive, Self-Optimizing Systems
34mOffline Reinforcement Learning and Model-Based Optimization
37mImplementing Reinforcement Learning for Marketing Channel Optimization with Python
30mPredictive Analytics in Social Media and Online Display Advertising
01h 04mProtective Optimization Technologies
47mAdvanced Reinforcement Learning & Data Science for Social Impact