Website CV Paper
Fields
Econometrics
Job market paper
Optimal batched contextual bandits under weak signals
When treatments are assigned over time, updating assignment rules using past observations can improve welfare. However, finding an optimal way to update these rules requires accounting for how current assignments affect future observations and decisions, making the problem numerically intractable as the number of individuals grows. We consider a simpler limiting problem based on weak-signal asymptotics and show that its solution is asymptotically optimal in the original problem. We then propose a new limiting problem robust to an unknown reward component shared across treatments. In simulations, a neural-network policy trained in this limiting problem achieves higher welfare than practical myopic learning algorithms such as Thompson sampling.