Unverified50% confidenceFactExact time
Algorithms such as online gradient descent can guarantee O(√T) regret under convex or approximately convex assumptions.
1
Sources
50%
Confidence
Long-term
Relevance
10/21/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedClassical results show that under convex losses, the cumulative regret of projected online gradient ascent is bounded by O(√T), where T is the number of rounds (Zinkevich, 2003).78% similarUnverifiedA good online learning algorithm should guarantee that regret grows far below linearly with the number of rounds, with a typical target on the order of O(√T); Exponential Weights (Hedge) and UCB have been proven to achieve this guarantee.75% similarUnverifiedIn standard regret minimization, the regret upper bound typically takes the form O(√T), whereas under the tilted sampling framework, sampling regret is dominated by the square of the confidence width, yielding logarithmic-level regret O(log T).74% similarUnverifiedThe distribution shift problem in behavior cloning causes error accumulation, with the expected total loss after T steps being O(εT²) rather than the linear O(εT).70% similarUnverifiedThe algorithm only needs the confidence width to shrink at a rate of 1/√t, and after accumulating the squares, a logarithmic regret is obtained.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1078996API
curl https://kongchang.com/api/v1/knowledge/claims/1078996MCP
get_claim(id=1078996)