Explore Once, Evaluate Many: Multiple Policy Evaluation for Stochastic Linear Bandits
PRISMA estimates multiple policies’ rewards from shared samples, with a sample bound that grows only logarithmically in the number of policies.
The guarantee applies to stochastic contextual bandits with a known linear reward feature map and reward-free access to policy queries. The algorithm combines sampling design with one shared ridge regression estimator to estimate every target policy’s value within a specified error, with high probability. Its guarantees require no numerical lower bound on second-moment eigenvalues and have no explicit dependence on the sizes of the context or action spaces. Policy queries are counted separately: each shared planning context requires one query per target policy.
ArXiv · AI/CL/LG's note
The guarantee applies to stochastic contextual bandits with a known linear reward feature map and reward-free access to policy queries. The algorithm combines sampling design with one shared ridge regression estimator to estimate every target policy’s value within a specified error, with high probability. Its guarantees require no numerical lower bound on second-moment eigenvalues and have no explicit dependence on the sizes of the context or action spaces. Policy queries are counted separately: each shared planning context requires one query per target policy.
ArXiv · AI/CL/LG's note
score 4