Panel Discussion
summary
Panelists argue that in the next five years batch/offline RL will move from demos to real-world proposals for policy design (e.g. tax or macro suggestions) and robotics, but only if methods address model mismatch and robustness rather than assuming realizable models. They highlight concrete tools: pessimism for off-policy RL to avoid extrapolation error, adaptive MPC runs a simulation and updates for closed‑loop safety, and off‑policy evaluation limits and variance as the bottleneck to reliable policy deployment.
AI-written summary · about
chapters (7)
- 0:00Start
- 4:32Optimizing the Tax Code
- 13:19Fundamental Limitations
- 23:30.How Do We Integrate the Simulation into Our Control Strategies in a Systematic Way
- 24:22Model Reduction and Learning
- 31:50How Do We Actually Do Decisions from Data
- 35:45Do We Need a Model on the System
related talks
19mTraining Agentic Reasoners
16mReal-World Robotics as the Next Frontier for AI?
46mEfficient Policy Optimization Techniques for LLMs
38mRetrospective on multi-agent learning as a path to intelligence
19mRL for Autonomous Coding
58mPanel Discussion: AI- Friend or Foe?
45mPanel Talk
06mRL in real life: Bringing reinforcement learning to the enterprise