Toronto, ON, Canada Nilesh Verma, Problem Setup sider a data stream {(xC,~C )}C 1 where xC 2 X and ~C 2 Y. C be a compact conguration space equipped with the metric ). Each conguration 2 2 C corresponds to a trained model with instantaneous loss ✓(52 (x),~). e dene the regret over a period ) as: '()) = ) ’ C=1 ✓(52C (xC ),~C ) ) ’ C=1 ✓(52⇤ C (xC ),~C ), e the instantaneous optimal conguration is given by: 2⇤ C = argmin 22C E ⇥ ✓(52 (xC ),~C ) ⇤ . By choosing the windo obtain a regret '()) that ) + log |C|, up to logarith '()) = ˜ $ T 4.2 (L B budget ) , there exists a se " ’ 9=1 3(2⇤ 9 ,2⇤ 9+1 ) 以下の -(/) を⼩さくするオンライン予測が「良いアルゴリズム」 理論的解析 研究会 © 2025 Naoki Chihara et al. 25 52 (·) with instantaneous loss ✓(52 (x),~). We dene the regret over a period ) as: '()) = ) ’ C=1 ✓(52C (xC ),~C ) ) ’ C=1 ✓(52⇤ C (xC ),~C ), where the instantaneous optimal conguration is given by: 2⇤ C = argmin 22C E ⇥ ✓(52 (xC ),~C ) ⇤ . 4.2 Assumptions (1) Bounded Loss: For all 2 2 C and (x,~), there exists ⌫ > 0 such that 0 ✓(52 (x),~) ⌫. (2) Lipschitz Continuity [14]: For all 2,20 2 C, there exists ! > 0 with T budget ) , th " ’ 9=1 3 P S intervals. Ea these yields ider a data stream {(xC,~C )}C 1 where xC 2 X and ~C 2 Y. be a compact conguration space equipped with the metric . Each conguration 2 2 C corresponds to a trained model with instantaneous loss ✓(52 (x),~). e dene the regret over a period ) as: '()) = ) ’ C=1 ✓(52C (xC ),~C ) ) ’ C=1 ✓(52⇤ C (xC ),~C ), e the instantaneous optimal conguration is given by: 2⇤ C = argmin 22C E ⇥ ✓(52 (xC ),~C ) ⇤ . Assumptions Bounded Loss: For all 2 2 C and (x,~), there exists ⌫ > 0 such that obtain a regret '( ) + log |C|, up to '() T 4.2 (L budget ) , there exi " ’ 9=1 3(2⇤ 9 ,2 P S. intervals. Each int these yields 52 (·) with instantaneous loss ✓(52 (x),~). We dene the regret over a period ) as: '()) = ) ’ C=1 ✓(52C (xC ),~C ) ) ’ C=1 ✓(52⇤ C (xC ), where the instantaneous optimal conguration is gi 2⇤ C = argmin 22C E ⇥ ✓(52 (xC ),~C ) ⇤ . 4.2 Assumptions (1) Bounded Loss: For all 2 2 C and (x,~), ther such that 0 ✓(52 (x),~) ⌫. (2) Lipschitz Continuity [14]: For all 2,20 2 C ! > 0 with BST が {"! }!"# $ を選んだ時の誤差の総和 バッチ処理で {"! }!"# $ を選んだ時の誤差の総和 系列全体を考慮して選んだ時刻 % における最良の "%