ORUCB
An online learning algorithm for deferral decision scenarios. It combines polynomial modeling of shared and expert-specific responses with confidence-weighted risk regression and exploration calibration, enabling correction of imperfect expert answers and optimization of routing strategies.
Timeline (last 90 days)
A study published on arXiv (arXiv:2610.07084) proposes the perspective of 'A Query Is Not a Commitment,' in which a learner must choose which expert to consult and pre-specify a correction function before paying to purchase the expert's answer.
The algorithm proposed in the study is named ORUCB, and its core idea is to pool shared responses with expert-specific responses through polynomial modeling.
ORUCB sets an upper bound on the cumulative response learning error and uses this bound to calibrate confidence-weighted risk regression and exploration strategies.
Under assumptions including bounded residuals and divergence, a fixed feasible model with optimal responses, free-query risk, and a linear model of optimal query risk, ORUCB can achieve a high-probability pseudo-regret bound of O(√T·log(T+1)) over T rounds.
The lower bound for the standard multi-armed bandit problem is O(√T), and ORUCB's extra log(T+1) factor stems from the complexity introduced by simultaneously learning the correction function.
The theoretical guarantees of ORUCB permit singular answer distributions, tolerate misspecified shared response models, and its optimality is relative to the class of bounded responses.
The study validates the method on four test data streams, and the cubic policy outperforms seven baseline methods that directly deploy uncorrected answers in terms of total cost including fees.
The experiments are conducted along two dimensions: comparing routing decisions against common correction learners, and comparing the relationship between cost and query rate across six pricing scenarios.
Pre-specifying the correction function and incorporating correction error into the routing decision yields considerable benefits at the economic cost level, especially when each query requires payment.
All Facts (9)
A study published on arXiv (arXiv:2610.07084) proposes the perspective of 'A Query Is Not a Commitment,' in which a learner must choose which expert to consult and pre-specify a correction function before paying to purchase the expert's answer.
50%UnverifiedThe algorithm proposed in the study is named ORUCB, and its core idea is to pool shared responses with expert-specific responses through polynomial modeling.
50%UnverifiedORUCB sets an upper bound on the cumulative response learning error and uses this bound to calibrate confidence-weighted risk regression and exploration strategies.
50%UnverifiedUnder assumptions including bounded residuals and divergence, a fixed feasible model with optimal responses, free-query risk, and a linear model of optimal query risk, ORUCB can achieve a high-probability pseudo-regret bound of O(√T·log(T+1)) over T rounds.
50%UnverifiedThe lower bound for the standard multi-armed bandit problem is O(√T), and ORUCB's extra log(T+1) factor stems from the complexity introduced by simultaneously learning the correction function.
50%UnverifiedThe theoretical guarantees of ORUCB permit singular answer distributions, tolerate misspecified shared response models, and its optimality is relative to the class of bounded responses.
50%UnverifiedThe study validates the method on four test data streams, and the cubic policy outperforms seven baseline methods that directly deploy uncorrected answers in terms of total cost including fees.
50%UnverifiedThe experiments are conducted along two dimensions: comparing routing decisions against common correction learners, and comparing the relationship between cost and query rate across six pricing scenarios.
50%UnverifiedPre-specifying the correction function and incorporating correction error into the routing decision yields considerable benefits at the economic cost level, especially when each query requires payment.
50%