NJUAT 南京大学 人工智能学院 SCHOOL OF ARTFICIAL INTELUGENCE,NANJING UNFVERSITY Lecture 9.Optimistic Online Mirror Descent Advanced Optimization(Fall 2023) Peng Zhao zhaop@lamda.nju.edu.cn Nanjing University
Lecture 9. Optimistic Online Mirror Descent Peng Zhao zhaop@lamda.nju.edu.cn Nanjing University Advanced Optimization (Fall 2023)
Beyond the Worst-Case Analysis All above regret guarantees hold against the worst case Matching the minimax optimality oblivious adversary adaptive adversary The environment is fully adversarial examination interview However,in practice: We are not always interested in the worst-case scenario Environments can exhibit specific patterns:gradual change,periodicity... We are after some more problem-dependent guarantees. Advanced Optimization(Fall 2023) Lecture 9.Optimistic Online Mirror Descent 2
Advanced Optimization (Fall 2023) Lecture 9. Optimistic Online Mirror Descent 2 Beyond the Worst-Case Analysis • All above regret guarantees hold against the worst case • Matching the minimax optimality • The environment is fully adversarial interview oblivious adversary adaptive adversary examination We are after some more problem-dependent guarantees. • However, in practice: • We are not always interested in the worst-case scenario • Environments can exhibit specific patterns: gradual change, periodicity…
Beyond the Worst-Case Analysis Beyond the worst-case analysis,achieving more adaptive results. (1)adaptivity:achieving better guarantees in easy problem instances; (2)robustness:maintaining the same worst-case guarantee. VS OPTIMISTS PESSIMISTS Be cautiously optimistic Real world Easy Data Worst-Case Data [Slides from Dylan Foster,Adaptive Online Learning @NIPS'15 workshop] Advanced Optimization(Fall 2023) Lecture 9.Optimistic Online Mirror Descent 3
Advanced Optimization (Fall 2023) Lecture 9. Optimistic Online Mirror Descent 3 Beyond the Worst-Case Analysis • Beyond the worst-case analysis, achieving more adaptive results. (1) adaptivity: achieving better guarantees in easy problem instances; (2) robustness: maintaining the same worst-case guarantee. [Slides from Dylan Foster, Adaptive Online Learning @NIPS’15 workshop]
Small-Loss Bounds for PEA Theorem2.Suppose that∀t∈[T]andi∈[N],0≤Lt,i≤l,then Hedge with learning rate nE(0,1)guarantees which can further imply t=1 7≤（，+）=o(vg+s by setting n=min When LT.i=O(T),it can recover the minimax O(VTlog N)guarantee. When Lr.=O(1),the regret bound is (log N),which is independent of T! Advanced Optimization(Fall 2023) Lecture 9.Optimistic Online Mirror Descent 4
Advanced Optimization (Fall 2023) Lecture 9. Optimistic Online Mirror Descent 4 Small-Loss Bounds for PEA
Small-Loss Bounds for PEA Addressing the unpleasant dependence on Lr.ivia self-confident tuning. Theorem4.Suppose that∀t∈[T]andi∈[N],0≤，i≤l,then Hedge with adaptive learning rate gurantees Regretr <8(Lr.+1)In N+3In N =O(V,1ogN+logN）, whereL is cumulative loss the learner suffered at time t. Advanced Optimization(Fall 2023) Lecture 9.Optimistic Online Mirror Descent 5
Advanced Optimization (Fall 2023) Lecture 9. Optimistic Online Mirror Descent 5 Small-Loss Bounds for PEA
Key Analysis in Self-confident Tuning Proof.From the potential-based proof,we already know that T T ∑m.-≤-1+mN+-p: t=1 ≤Vr-1+nN+∑ P,) 1=V V∑ip,)+1 (L-1-∑号p》 How to bound this term? This is actually a common structure to handle. Advanced Optimization(Fall 2023) Lecture 9.Optimistic Online Mirror Descent 6
Advanced Optimization (Fall 2023) Lecture 9. Optimistic Online Mirror Descent 6 Key Analysis in Self-confident Tuning Proof. From the potential-based proof, we already know that How to bound this term? This is actually a common structure to handle
Small-Loss Bound for PEA:Proof Proof.From previous potential-based proof,we already known that 含m-sv+m同 Pe) Lemma 2.Let a1,a2,...,ar be non-negative real numbers.Then T T te T] Lr-Lr,≤V(亿r-1+1)lnN+4y1+Lr+1 (C,≤L∈N) <V(Lr+1)In N+4v1+Lr+1 Advanced Optimization(Fall 2023) Lecture 9.Optimistic Online Mirror Descent 7
Advanced Optimization (Fall 2023) Lecture 9. Optimistic Online Mirror Descent 7 Small-Loss Bound for PEA: Proof Proof. From previous potential-based proof, we already known that
Small-Loss Bounds for OCO Definition 4(Small Loss).The small-loss quantity of the OCO problem(online function f:R)is defined as T Fr=min x∈X fix) t=1 One essential property for small-loss bound for OCO:self-bounding property. Corollary 1.For an L-smooth and non-negative function f:RdR,we have that IVf(x)2≤V2Lf(x),x∈X. Advanced Optimization(Fall 2023) Lecture 9.Optimistic Online Mirror Descent 8
Advanced Optimization (Fall 2023) Lecture 9. Optimistic Online Mirror Descent 8 Small-Loss Bounds for OCO One essential property for small-loss bound for OCO: self-bounding property
Achieving Small-Loss Bound We show that under the self-bounding condition,OGD can yield the desired small-loss regret bound. xt+1=Πx∈xXt-:Vf(xt)J Theorem 6(Small-loss Bound).Assume that fr is L-smooth and non-negative for all t E when settingthe regret of OGD to any comparator is bounded as Regretr=∑ix)-∑fm≤o(V+同） where Gis the empirical cumulative gradient norm. Advanced Optimization(Fall 2023) Lecture 9.Optimistic Online Mirror Descent 9
Advanced Optimization (Fall 2023) Lecture 9. Optimistic Online Mirror Descent 9 Achieving Small-Loss Bound • We show that under the self-bounding condition, OGD can yield the desired small-loss regret bound
Proof Pof)-m≤ix+2品(u--u- 玄u度 +≤01+空+ (11) (G=∑1Vf(x)) Lemma 1.Let a,a2,...,ar be non-negative real numbers.Then T T -≤1+ at t=11 Advanced Optimization(Fall 2023) Lecture 9.Optimistic Online Mirror Descent 10
Advanced Optimization (Fall 2023) Lecture 9. Optimistic Online Mirror Descent 10 Proof Proof