t = 1, . . . , T : 1. 2. 環境 3. 凸集合 K ⊂ Rd 行動 xt 選択(K : 実行可能領域) 凸損失関数 ft : K → R 決定( Lipschitz 連続) 損失 ft (xt ) 目標: 被 RegT RegT = T X ,gt ∈ ∂ft (xt )( ft ) 観測 最小化 ft (xt ) t=1 累積損失 − min x∈K 最適 T X ft (x) t=1 固定点 累積損失
t = 1, . . . , T : 凸集合 K ⊂ Rd 行動 xt 選択(K : 実行可能領域) 凸損失関数 ft : K → R 決定( Lipschitz 連続) 1. 2. 環境 損失 ft (xt ) 3. 目標: 被 ,gt ∈ ∂ft (xt )( 最小化 RegT RegT = T X − ft (xt ) t=1 累積損失 • 損失関数 ft 問題 • 標準的 ft ) 観測 gt ∈ R d 用 線形最適化 √ RegT = O( T ) min x∈K T X ft (x) t=1 最適 ft (·) = ⟨gt , ·⟩ 呼 (勾配降下法 固定点 累積損失 , )
対 ,F (w ) = Eξ [f (w , ξ)] 例. 二乗損失 用 線形回帰: f (w , (z, y ) ) = (⟨w , z⟩ − y )2 | {z } 最小値 求 問題 入力 ・ 定理 対 動 変換 (Cesa-Bianchi et al., 2004) 凸最適化 , " Eξ1 ,...,ξT F 損失関数 ft (w ) := f (w , ξt ) (ξ1 , . . . , ξT ∼ ρ) 1 X wt T T t=1 !# − F (w ∗ ) ≤ Eξ1 ,...,ξT [RegT ] . T 例.確率的勾配降下法 (Zinkevich, 2003),AdaGrad (McMahan and Streeter, 2010; Duchi et al., 2011)
対 ,F (w ) = Eξ [f (w , ξ)] 例. 二乗損失 用 線形回帰: f (w , (z, y ) ) = (⟨w , z⟩ − y )2 | {z } 最小値 求 問題 入力 ・ 定理 対 動 変換 (Cesa-Bianchi et al., 2004) 凸最適化 , " Eξ1 ,...,ξT F 損失関数 ft (w ) := f (w , ξt ) (ξ1 , . . . , ξT ∼ ρ) 1 X wt T T t=1 !# − F (w ∗ ) ≤ Eξ1 ,...,ξT [RegT ] . T 例.確率的勾配降下法 (Zinkevich, 2003),AdaGrad (McMahan and Streeter, 2010; Duchi et al., 2011) 証明( 場合). P Jensen P RegT T 1 ∗ f T t=1 wt − f (w ∗ ) ≤ T1 T t=1 (f (wt ) − f (w )) ≤ T .
F 1 X wt T T !# − F (w ∗ ) ≤ t=1 直接的 適用 √ • f 凸関数:RegT = O( T ) → 収束 • f µ-強凸: RegT = O( µ1 log T ) → 収束 Eξ1 ,...,ξT [RegT ] . T √ O(1/ T ) O(log(T )/(µT )) 非最適
s0 ◦:状態 s ∈ S →:行動 a ∈ A (遷移確率 P At each episode t = 1, . . . , T : 1. 環境 損失関数 ℓt : S × A → [0, 1] 決定 2. 方策 πt : S → ∆(A)(状態 3. s0 πt 従 行動 行動 ,訪 確率分布) 決定 (s , a ) 損失 観測 遷移 既知)
) s2 Markov 決定過程上 学習 s3 s0 ℓt (s0 , a0 ) s1 ℓt (s1 , a1 ) ◦:状態 s ∈ S →:行動 a ∈ A (遷移確率 P ··· 遷移 既知) At each episode t = 1, . . . , T : 1. 環境 損失関数 ℓt : S × A → [0, 1] 決定 2. 方策 πt : S → ∆(A)(状態 3. s0 πt 目標: 従 行動 行動 ,訪 (s , a ) Reg 最小化 " T T # X E V πt (ℓt ) − RegT = min E π t=1 累積損失 V π (ℓ):損失 ℓ s0 方策 π 従 確率分布) 決定 最適 1 損失 " T X 観測 # V π (ℓt ) t=1 固定方策 累積損失 期待累積損失(価値関数)
環境下 累積損失 最小化 問題 At each round t = 1, . . . , T : 1. 2. 環境 3. 凸集合 K ⊂ Rd xt 凸損失関数 ft : K → R 損失 ft (xt ) 目標: 決定( Lipschitz 連続) ,gt ∈ ∂ft (xt )( 被 T X ft (xt ) − t=1 累積損失 損失関数 ft 問題 gt ∈ R d 用 線形最適化 ft ) 観測 最小化 RegT RegT = 選択(K : 実行可能領域) ft (·) = ⟨gt , ·⟩ 呼 . min x∈K 最適 T X ft (x) t=1 固定点 累積損失 ,
(x, y ) = ψ(x) − (ψ(y ) + ⟨∇ψ(y ), x − y ⟩) ψ(z) 例: 1. ψ(x) = 12 ∥x∥2 → Dψ (x, y ) = 12 ∥x − y ∥2 P 2. ψ(x) = di=1 x(i) log x(i), x ∈ ∆d (negative) Shannon → Dψ (x, y ) = d X i=1 x(i) x(i) log y (i) KL =: KL(x, y ) Dψ (x, y ) ψ(y ) + ⟨∇ψ(y ), z − y ⟩ y x z
− η gt ) ⇐⇒ n o 1 xt+1 = arg minx∈K ⟨gt , x ⟩ + 2η ∥x − xt ∥2 ∥x1 − u ∥2 ηX + ∥gt ∥2identity matrix 2η 2 T RegT (u ) ≤ t=1 Hedge : P xt+1 (i ) ∝ exp −η ts=1 gs (i ) ⇐⇒ n o xt+1 = arg minx∈∆d ⟨gt , x ⟩ + η1 KL(x , xt ) XX X KL(ei ∗ , x1 ) log d +η x t (i )g t (i )2 = +η ∥gt ∥2(∇2 −H(xt ))−1 η η T RegT (ei ∗ ) ≲ ∀ i ∈ [d ] d t=1 i=1 T t=1 (∥x∥A = √ x ⊤ Ax)
− η gt ) ⇐⇒ n o 1 xt+1 = arg minx∈K ⟨gt , x ⟩ + 2η ∥x − xt ∥2 ∥x1 − u ∥2 ηX + ∥gt ∥2identity matrix 2η 2 T RegT (u ) ≤ t=1 : Hedge P xt+1 (i ) ∝ exp −η ts=1 gs (i ) ⇐⇒ n o xt+1 = arg minx∈∆d ⟨gt , x ⟩ + η1 KL(x , xt ) XX X KL(ei ∗ , x1 ) log d +η x t (i )g t (i )2 = +η ∥gt ∥2(∇2 −H(xt ))−1 η η T RegT (ei ∗ ) ≲ ∀ i ∈ [d ] d T t=1 i=1 ??? xt+1 = arg min ⟨gt , x ⟩ + ?? =⇒ √ t=1 : x∈K (∥x∥A = 1 Dψ ( x , x t ) , η Dψ (x , y ) = ψ(x ) − ψ(y ) − ⟨∇ψ(y ), x − y ⟩ ?? Dψ (u , x1 ) RegT (u ) ≤ η +η T X t=1 ∥gt ∥2(∇2 ψ(xt ))−1 x ⊤ Ax)
(t) 定理 (Freund and Schapire, 1999) 戦略 平均 x̄T = T1 t=1 x ,ȳT = T t=1 y T T 直積分布 (x̄T , ȳT ) ((Regx + Regy )/T )-近似 Nash 均衡 証明.以下 ⟨x, AȳT ⟩ − 示 十分 T RegT x + Regy T ≤ ⟨x̄T , AȳT ⟩ ≤ ⟨x̄T , Ay ⟩ + T RegT x + Regy T ∀x ∈ ∆mx , y ∈ ∆my
(t) 定理 (Freund and Schapire, 1999) 戦略 平均 x̄T = T1 t=1 x ,ȳT = T t=1 y T T 直積分布 (x̄T , ȳT ) ((Regx + Regy )/T )-近似 Nash 均衡 証明. 定義 , X 1 X 1 ⟨x̄T , AȳT ⟩ = ⟨x̄T , Ay (t) ⟩ ≤ max ⟨x ∗ , Ay (t) ⟩ T T x∗ T T t=1 t=1
(t) 定理 (Freund and Schapire, 1999) 戦略 平均 x̄T = T1 t=1 x ,ȳT = T t=1 y T T 直積分布 (x̄T , ȳT ) ((Regx + Regy )/T )-近似 Nash 均衡 証明. 定義 , X 1 X 1 ⟨x̄T , AȳT ⟩ = ⟨x̄T , Ay (t) ⟩ ≤ max ⟨x ∗ , Ay (t) ⟩ T T x∗ T T t=1 RegTx = maxx ∗ PT ∗ (t) ⟩− t=1 ⟨x , Ay PT (t) , Ay (t) ⟩ t=1 ⟨x t=1 = 1 T T X t=1 ⟨x (t) , Ay (t) ⟩ + RegT x T
(t) 定理 (Freund and Schapire, 1999) 戦略 平均 x̄T = T1 t=1 x ,ȳT = T t=1 y T T 直積分布 (x̄T , ȳT ) ((Regx + Regy )/T )-近似 Nash 均衡 証明. 定義 , X 1 X 1 ⟨x̄T , AȳT ⟩ = ⟨x̄T , Ay (t) ⟩ ≤ max ⟨x ∗ , Ay (t) ⟩ T T x∗ T T t=1 RegTx = maxx ∗ PT ∗ (t) ⟩− t=1 ⟨x , Ay PT (t) , Ay (t) ⟩ t=1 ⟨x P P RegTy = Tt=1 ⟨y (t) , A⊤ x (t) ⟩ − miny ∗ Tt=1 ⟨y ∗ , A⊤ x (t) ⟩ t=1 = 1 T T X ⟨x (t) , Ay (t) ⟩ + t=1 RegT x T T X RegxT + RegT 1 y (t) ∗ ⟨x , Ay ⟩ + = min T y∗ T t=1
(t) 定理 (Freund and Schapire, 1999) 戦略 平均 x̄T = T1 t=1 x ,ȳT = T t=1 y T T 直積分布 (x̄T , ȳT ) ((Regx + Regy )/T )-近似 Nash 均衡 証明. 定義 , X 1 X 1 ⟨x̄T , AȳT ⟩ = ⟨x̄T , Ay (t) ⟩ ≤ max ⟨x ∗ , Ay (t) ⟩ T T x∗ T T t=1 RegTx = maxx ∗ PT ∗ (t) ⟩− t=1 ⟨x , Ay t=1 PT (t) , Ay (t) ⟩ t=1 ⟨x P P RegTy = Tt=1 ⟨y (t) , A⊤ x (t) ⟩ − miny ∗ Tt=1 ⟨y ∗ , A⊤ x (t) ⟩ = 1 T T X ⟨x (t) , Ay (t) ⟩ + t=1 RegT x T T X RegxT + RegT 1 y (t) ∗ ⟨x , Ay ⟩ + = min T y∗ T t=1 x̄T 定義 = min ⟨x̄T , Ay ∗ ⟩ + ∗ y T RegT x + Regy . T
解析 RegT x = ∗max x ∈∆mx ≤ (細 ( T X ⟨x , g t=1 T X log mx +η η ∗ t=1 (t) ⟩− T X ) ⟨x (t) ,g (t) ⟩ t=1 1 X (t+1) ∥x − x (t) ∥21 4η T ∥g (t) − g (t−1) ∥2∞ − optimistic t=1
解析 RegT x = ∗max x ∈∆mx ≤ (細 ( T X ⟨x , g t=1 T X log mx +η η ∗ t=1 (t) ⟩− T X ) ⟨x (t) ,g (t) ⟩ t=1 1 X (t+1) ∥x − x (t) ∥21 4η T ∥g (t) − g (t−1) ∥2∞ − optimistic t=1 第2項 ∥g (t) − g (t−1) ∥2∞ = ∥A(y (t) − y (t−1) )∥2∞ = max |⟨A(k , ·), y (t) − y (t−1) ⟩|2 ≤ ∥y (t) − y (t−1) ∥21 . k∈[mx ]
解析 RegT x = ∗max x ∈∆mx ≤ (細 ( T X ∗ ⟨x , g t=1 T X log mx +η η (t) ⟩− T X ) ⟨x (t) ,g (t) ⟩ t=1 1 X (t+1) ∥x − x (t) ∥21 4η T ∥g (t) − g (t−1) ∥2∞ − t=1 optimistic t=1 第2項 ∥g (t) − g (t−1) ∥2∞ = ∥A(y (t) − y (t−1) )∥2∞ = max |⟨A(k , ·), y (t) − y (t−1) ⟩|2 ≤ ∥y (t) − y (t−1) ∥21 . k∈[mx ] 2 不等式 合 RegT x ≤ X (t) log mx 1 X (t+1) +η ∥y − y (t−1) ∥21 − ∥x − x (t) ∥21 η 4η T T t=1 t=1
(t) log mx 1 X (t+1) +η ∥y − y (t−1) ∥21 − ∥x − x (t) ∥21 . η 4η (★-x) X (t) log my 1 X (t+1) +η ∥x − x (t−1) ∥21 − ∥y − y (t) ∥21 . η 4η (★-y) T T t=1 t=1 同様 RegT y ≤ T T t=1 t=1
(t) log mx 1 X (t+1) +η ∥y − y (t−1) ∥21 − ∥x − x (t) ∥21 . η 4η (★-x) X (t) log my 1 X (t+1) +η ∥x − x (t−1) ∥21 − ∥y − y (t) ∥21 . η 4η (★-y) T T t=1 t=1 同様 RegT y ≤ 2 不等式 足 , log(mx my ) T RegT + x + Regy ≤ η T T t=1 t=1 1 η− 4η | {z } =−3/4(η = 1/4 T X t=1 ) ∥x (t) − x (t−1) ∥21 + T X t=1 ! ∥y (t) − y (t−1) ∥21
(t) log mx 1 X (t+1) +η ∥y − y (t−1) ∥21 − ∥x − x (t) ∥21 . η 4η (★-x) X (t) log my 1 X (t+1) +η ∥x − x (t−1) ∥21 − ∥y − y (t) ∥21 . η 4η (★-y) T T t=1 t=1 同様 RegT y ≤ 2 不等式 足 , T T t=1 t=1 1 η− 4η | {z } log(mx my ) T RegT + x + Regy ≤ η T X t=1 =−3/4(η = 1/4 η = 1/4 (η − 1/(4η) < 0 ) ), ∥x (t) − x (t−1) ∥21 + T X t=1 ! ∥y (t) − y (t−1) ∥21
(t) log mx 1 X (t+1) +η ∥y − y (t−1) ∥21 − ∥x − x (t) ∥21 . η 4η (★-x) X (t) log my 1 X (t+1) +η ∥x − x (t−1) ∥21 − ∥y − y (t) ∥21 . η 4η (★-y) T T t=1 t=1 同様 RegT y ≤ 2 不等式 足 , T T t=1 t=1 1 η− 4η | {z } log(mx my ) T RegT + x + Regy ≤ η T X t=1 =−3/4(η = 1/4 η = 1/4 (η − 1/(4η) < 0 ∥x (t) − x (t−1) ∥21 + T X ! ∥y (t) − y (t−1) ∥21 t=1 ) ), T RegT x + Regy ≥ 0 注意 ,
(t) log mx 1 X (t+1) +η ∥y − y (t−1) ∥21 − ∥x − x (t) ∥21 . η 4η (★-x) X (t) log my 1 X (t+1) +η ∥x − x (t−1) ∥21 − ∥y − y (t) ∥21 . η 4η (★-y) T T t=1 t=1 同様 RegT y ≤ 2 不等式 足 , T T t=1 t=1 1 η− 4η | {z } log(mx my ) T RegT + x + Regy ≤ η T X t=1 =−3/4(η = 1/4 η = 1/4 (η − 1/(4η) < 0 T X t=1 ∥x (t) − x (t−1) ∥21 + t=1 − x (t−1) ∥21 + T X ! ∥y (t) − y (t−1) ∥21 t=1 ) ), T X ∥x (t) T RegT x + Regy ≥ 0 ∥y (t) − y (t−1) ∥21 ≤ O (log(mx my )) . 注意 ,
(t) log mx 1 X (t+1) +η ∥y − y (t−1) ∥21 − ∥x − x (t) ∥21 . η 4η (★-x) X (t) log my 1 X (t+1) +η ∥x − x (t−1) ∥21 − ∥y − y (t) ∥21 . η 4η (★-y) T T t=1 t=1 同様 RegT y ≤ 不等式 2 足 , T T t=1 t=1 1 η− 4η | {z } log(mx my ) T RegT + x + Regy ≤ η T X t=1 =−3/4(η = 1/4 η = 1/4 (η − 1/(4η) < 0 T X ∥x (t) − x (t−1) ∥21 + t=1 (★-x),(★-y) − x (t−1) ∥21 + T X ! ∥y (t) − y (t−1) ∥21 t=1 ) ), T X ∥x (t) T RegT x + Regy ≥ 0 注意 , ∥y (t) − y (t−1) ∥21 ≤ O (log(mx my )) . t=1 合 T ,RegT x = O (log(mx my )),Regy = O (log(mx my )) 得 .
e (·):T 以外 対数 factor O 行動数,T : 数, 省略 文献 上界 Blum and Mansour (2007) Chen and Peng (2020) Anagnostides et al. (2022a) Anagnostides et al. (2022b) → 還元 現 最小化:行動数分 Markov 連鎖 定常分布 e (√mT log m) O e (√n m3/4 T 1/4 ) O e (nm4 log4 T ) O O (nm5/2 log T ) Hedge log-barrier → 定常分布 安定性制御 最小化 還元 安定性 制御 (Appendix 参照)
e (·):T 以外 対数 factor O 行動数,T : 数, 省略 文献 上界 Blum and Mansour (2007) Chen and Peng (2020) Anagnostides et al. (2022a) Anagnostides et al. (2022b) Tsuchiya (2026) → 還元 現 最小化:行動数分 Markov 連鎖 定常分布 e (√mT log m) O e (√n m3/4 T 1/4 ) O e (nm4 log4 T ) O O (nm5/2 log T ) e (nm2 √log T ) O Hedge log-barrier → 定常分布 安定性制御 log-barrier → 定常分布 entropy → 学習器 最小化 還元 安定性 制御 (Appendix 参照)
e (·):T 以外 対数 factor O 行動数,T : 数, 省略 文献 上界 Blum and Mansour (2007) Chen and Peng (2020) Anagnostides et al. (2022a) Anagnostides et al. (2022b) Tsuchiya (2026) Mai (2026) → 還元 現 最小化:行動数分 Markov 連鎖 定常分布 e (√mT log m) O e (√n m3/4 T 1/4 ) O e (nm4 log4 T ) O Hedge log-barrier → 定常分布 安定性制御 O (nm5/2 log T ) log-barrier → 定常分布 e (nm2 √log T ) O entropy → 学習器 √ e O ( n m) (T 非依存) 最小化 還元 安定性 制御 (Appendix 参照)
Ashok Cutkosky (2025). “General framework for online-to-nonconvex conversion: Schedule-free SGD is also effective for nonconvex optimization”. In: Proceedings of the 42nd International Conference on Machine Learning. Vol. 267, pp. 772–795. Kwangjun Ahn et al. (2024). “Understanding Adam optimizer via online learning of updates: Adam is FTRL in disguise”. In: Proceedings of the 41st International Conference on Machine Learning. Vol. 235, pp. 619–640. Ioannis Anagnostides et al. (2022a). “Near-optimal no-regret learning for correlated equilibria in multi-player general-sum games”. In: Proceedings of the 54th Annual ACM SIGACT Symposium on Theory of Computing, pp. 736–749. Ioannis Anagnostides et al. (2022b). “Uncoupled Learning Dynamics with O(log T ) Swap Regret in Multiplayer Games”. In: Advances in Neural Information Processing Systems. Vol. 35, pp. 3292–3304. Robert J. Aumann (1974). “Subjectivity and correlation in randomized strategies”. In: Journal of Mathematical Economics 1.1, pp. 67–96. James P. Bailey and Georgios Piliouras (2018). “Multiplicative Weights Update in Zero-Sum Games”. In: Proceedings of the 2018 ACM Conference on Economics and Computation, pp. 321–338. Avrim Blum and Yishay Mansour (2007). “From External to Internal Regret”. In: Journal of Machine Learning Research 8.47, pp. 1307–1324.
(2018). “Superhuman AI for heads-up no-limit poker: Libratus beats top professionals”. In: Science 359.6374, pp. 418–424. Nicolò Cesa-Bianchi, Alex Conconi, and Claudio Gentile (2004). “On the Generalization Ability of On-Line Learning Algorithms”. In: IEEE Transactions on Information Theory 50.9, pp. 2050–2057. Xi Chen and Binghui Peng (2020). “Hedging in games: Faster convergence of external and swap regrets”. In: Advances in Neural Information Processing Systems. Vol. 33, pp. 18990–18999. Ashok Cutkosky, Harsh Mehta, and Francesco Orabona (2023). “Optimal stochastic non-smooth non-convex optimization through online-to-non-convex conversion”. In: Proceedings of the 40th International Conference on Machine Learning. Vol. 202, pp. 6643–6670. Constantinos Daskalakis, Alan Deckelbaum, and Anthony Kim (2011). “Near-optimal no-regret algorithms for zero-sum games”. In: Proceedings of the Twenty-Second Annual ACM-SIAM Symposium on Discrete Algorithms, pp. 235–254. John Duchi, Elad Hazan, and Yoram Singer (2011). “Adaptive Subgradient Methods for Online Learning and Stochastic Optimization”. In: Journal of Machine Learning Research 12.61, pp. 2121–2159. Dean P. Foster and Rakesh V. Vohra (1997). “Calibrated Learning and Correlated Equilibrium”. In: Games and Economic Behavior 21.1–2, pp. 40–55.
Schapire (1999). “Adaptive Game Playing Using Multiplicative Weights”. In: Games and Economic Behavior 29.1–2, pp. 79–103. Chi Jin et al. (2020). “Learning Adversarial Markov Decision Processes with Bandit Feedback and Unknown Transition”. In: Proceedings of the 37th International Conference on Machine Learning. Vol. 119, pp. 4860–4869. Mingyi Li and Taira Tsuchiya (2026). “Muon with Finite Newton–Schulz: The Smoothing Benefit in Nonsmooth Nonconvex Optimization”. In: arXiv preprint arXiv:2608.26288. Haipeng Luo, Chen-Yu Wei, and Chung-Wei Lee (2021). “Policy Optimization in Adversarial MDPs: Improved Exploration via Dilated Bonuses”. In: Advances in Neural Information Processing Systems. Vol. 34, pp. 22931–22942. Tung Mai (2026). “Constant Swap Regret in General-Sum Games via Two-Scale Higher-Order Optimism”. In: arXiv preprint arXiv:2609.16751. H. Brendan McMahan and Matthew J. Streeter (2010). “Adaptive Bound Optimization for Online Convex Optimization”. In: Proceedings of the 23rd Conference on Learning Theory, pp. 244–256. Sasha Rakhlin and Karthik Sridharan (2013). “Optimization, Learning, and Games with Predictable Sequences”. In: Advances in Neural Information Processing Systems. Vol. 26, pp. 3066–3074.
“Optimistic Policy Optimization with Bandit Feedback”. In: Proceedings of the 37th International Conference on Machine Learning. Vol. 119, pp. 8604–8613. Vasilis Syrgkanis et al. (2015). “Fast Convergence of Regularized Learning in Games”. In: Advances in Neural Information Processing Systems. Vol. 28, pp. 2989–2997. Taira Tsuchiya (2026). “Sublogarithmic Swap Regret in Multiplayer General-Sum Games via Hybrid Regularization”. In: arXiv preprint arXiv:2608.04149. Chen-Yu Wei et al. (2021). “Linear Last-iterate Convergence in Constrained Saddle-point Optimization”. In: International Conference on Learning Representations. Qinzi Zhang and Ashok Cutkosky (2024). “Random scaling and momentum for non-smooth non-convex optimization”. In: Proceedings of the 41st International Conference on Machine Learning. Vol. 235, pp. 58780–58799. Alexander Zimin and Gergely Neu (2013). “Online learning in episodic Markovian decision processes by relative entropy policy search”. In: Advances in Neural Information Processing Systems. Vol. 26, pp. 1583–1591. Martin Zinkevich (2003). “Online Convex Programming and Generalized Infinitesimal Gradient Ascent”. In: Proceedings of the 20th International Conference on Machine Learning, pp. 928–935.