Scool (@InriaScool) profile picture

Scool is a #MachineLearning research team in @Inria & CRIStAL interested in designing algorithms that learn & adapt on-the-go. It is the new avatar of "SequeL".

Lille, France
Joined November 2020
4/n Alena, Thomas, Phillipe & Bruno motivated by uniqueness ambiguity of value func as a solution to HJB eqn in CTRL, propose to approximate the value func by training a PINN through a specific scheduling iterative process that constraints it to converge to the viscosity solution
2
141
3/n MCTS with deep NN shows promising performance in deterministic envs, but fails in stochastic envs. @tuanquangdam, Odarlic & Brahim propose CATS & PATS, leveraging TS to handle selection randomness. They achieve regret guarantees as well as good performance in stochastic envs.
1
2
217
2/n Hardness of pure exploration with known constraints changes with the geometry of the feasible set induced by constraints. Udvas & @BasuDebabrota propose LATS & LAGEX to tackle pure exp under unknown constraints. Their Lagrangian scheme generalises several existing problems.
1
1
3
31
1/n Today, we concluded @icmlconf with 4 presentations at the #FORLAC workshop conjoining RL theory and Control. Following their UAI work, @tuanquangdam, Odalric & Emilie on their work to address biased value function estimation in #MCTS using power means. #ICML2024 @Inria_Lille
1
3
396
We've derived tight lower and upper bounds for differentially private finite-armed & linear bandits, while we lack the same for contextual bandits. At #COLT2024, @achraf_azize presents open problems in contextual bandits with privacy. @BasuDebabrota @Inria_Lille @RechercheUlille
1
2
8
922
2/2 If each arm has multiple objectives, how to identify an arm whose mean vector is not worse than any of the others. Tomorrow @aistats_conf, Emilie & Cyrille will present "first" algo to detect such pareto sets with finite budget & bandit feedback. @RechercheUlille @Inria_Lille
3
73
1/2 Is exploration harder if we've constraints on policies? No, depends on how constraints change the geometry of alternating set. Today @aistats_conf, @BasuDebabrota & collaborators present insights & algorithms for pure exploration with constraints. #AISTATS2024 @chalmersuniv
1
6
183
2/2 Today @satml_conf, @achraf_azize is presenting these nuances of "Concentrated DP for Bandits" along with information-theoretic lower bounds and near-optimal algorithms for linear, contextual, and multi-armed bandits. #privacy #bandits @Inria_Lille @RechercheUlille #SaTML24
1
3
163
1/2 To define privacy in bandits, we have to ask what are the input and output of a bandit algorithm? What differs if the adversary is interactive or passive? @achraf_azize & @BasuDebabrota address these in their work openreview.net/forum?id=2366….
1
1
1
4
461
@TmlrOrg we address the corrupted bandit problem, i.e. a stochastic multi-armed bandit problem with unknown reward distributions, which are heavy-tailed and corrupted by a history-independent adversary or Nature. We provide another set of lower bounds and algorithm. #robustness
3
75
Today at #ALT2024, @ShubhadaAgrawal presents CRIMED: a joint work with Timothée, @BasuDebabrota & Odalric. CRIMED achieves matching regret upper bound for symmetric distributions and unbounded corruption. #bandits #corrupted_observations @univ_lille @Inria_Lille
1
4
102
What happens in a bandit problem if epsilon fraction of feedback are arbitrarily corrupt? What are the new lower bounds on the regret? Can we design an optimal algorithm for #Bandits_corrupted_by_nature? We address this question in two parts.
1
1
4
429