KL-UCB-Transfer
KL-UCB-Transfer is a bandit algorithm that incorporates transfer learning into the KL-UCB framework, achieving logarithmic regret even when historical data originate from an unknown adaptive policy. It benefits fully when the historical data are useful (obtaining a better leading constant) and automatically degrades to the guarantees of classical KL-UCB when the historical data are unreliable, exhibiting graceful degradation.
Timeline (last 90 days)
The researchers proposed the KL-UCB-Transfer algorithm, which combines online data with historical observations, extends the classical KL-UCB framework, and is able to achieve logarithmic regret.
The KL-UCB-Transfer algorithm obtains a better leading constant under the condition μ'_a < μ_1 - L_a, and degrades to the guarantee of classical KL-UCB when this condition is not satisfied.
All Facts (2)
The KL-UCB-Transfer algorithm obtains a better leading constant under the condition μ'_a < μ_1 - L_a, and degrades to the guarantee of classical KL-UCB when this condition is not satisfied.
50%UnverifiedThe researchers proposed the KL-UCB-Transfer algorithm, which combines online data with historical observations, extends the classical KL-UCB framework, and is able to achieve logarithmic regret.
50%