A Regret Minimization Approach to Mutli-Agent Control and RL
summary
Claim: multi-agent control reduces to distributed online convex optimization by linearize global cost into local gradients and letting each actuator run a no‑regret gradient method on that linearized loss. Tools: online convex optimization, regret minimization, and the disturbance-action controller policy class with gradient‑based updates. Result: provable global regret bounds (sum of local regrets plus low‑order term) and simulation on a four‑actuator airplane where the method outperforms baselines after a single actuator fails.
AI-written summary · about
chapters (12)
- 0:00Introduction to speaker
- 1:08Multi-agent control overview
- 4:46Online convex optimization
- 7:44Regret and optimality
- 10:48Linearization method
- 12:46Control system dynamics
- 15:47Robust vs online control
- 21:10Disturbance action control
- 23:40Multi-agent control model
- 25:44Policy classes and extensions
- 29:45Experimental results
- 31:15Q&A and closing
related talks
41mCompetitive Algorithms for Online Control
01h 17mOnline Reinforcement Learning and Regret
15mMulti-Agent Reinforcement Learning In Stochastic Games: From Alphago To Robust Control
41mBetter Learning from the Past: Counterfactual / Batch RL
48mA Simple Condition for Constant Regret in Online Decision-Making
31mLearning to Control Unknown Systems
31mRegret Minimization for Stochastic Shortest Paths
56mA Tutorial on Reinforcement Learning II