A & O Apprentissage & Optimisation Head: Mich` ele Sebag - - PowerPoint PPT Presentation

A & O Apprentissage & Optimisation

Head: Mich` ele Sebag Joint INRIA project, Head: Marc Schoenauer

1 / 22

Members – 2008-2013

Permanent members

◮

Jamal Atif, MdC IUT Orsay ← 2011

◮

Anne Auger, CR INRIA

◮

Nicolas Bred` eche, MdC Univ. Paris Sud → 2012

◮

Philippe Caillou, MdC IUT Sceaux

◮

Marta Franova, CR CNRS

◮

Cyril Furtlehner, CR INRIA

◮

C´ ecile Germain-Renaud, Prof. UPS

◮

Nikolaus Hansen, CR INRIA ← 2009

◮

Yann Ollivier, CR CNRS ← 2010

◮

Marc Schoenauer, DR INRIA

◮

Mich` ele Sebag, DR CNRS

◮

Nicolas Spyratos, Prof. UPS, emeritus ← 2012

◮

Olivier Teytaud, CR INRIA Associate members

◮

Florence d’Alch´ e-Buc, Pr. Evry

◮

Guillaume Charpiat, CR INRIA, Sophia

◮

Bal´ azs K´ egl, DR CNRS, LAL

◮

H´ el` ene Paugam-Moisy, Pr. Lyon-2

◮

R´ emi Peyre, MdC Nancy

University CNRS INRIA

Non-permanent members

◮ 14 PhDs, 4 post-docs, 3 research

engineers Some figures

◮ 18 PhDs and 3 HdRs defended ◮ Articles 39 / 7 ◮ Conferences 110 / 188

2 / 22

Apprentissage & Optimisation

Research goals Model, predict and control physical or artificial systems A unified perspective:

◮ Learning is an optimization problem ◮ Ill-posed optimisation requires adaptation

3 / 22

Structure

Core expertise

◮ Machine Learning ◮ Stochastic

Optimization

◮ Statistical Physics

Applications

◮ Games and Energy management ◮ Numerical engineering ◮ Autonomic Computing ◮ Robotics

Special Interest Groups

◮ Stochastic continuous optimization ◮ Optimal decision making under uncertainty ◮ Criteria design ◮ Algorithm control and parameter tuning ◮ Large scale modelling

4 / 22

SIG Stochastic continuous optimization Argmax F : Ω ⊂ I Rd → I R

Covariance-Matrix Adaptation, Evolution Strategy

◮ Invariances under monotonous transform of F and affine

• transf. of Ω.

◮ A particular case of Information Geometry Optimization

Ollivier, Auger, Hansen, Arnold, ArXiv

Transfert 100+ applications in industry OMD, EADS, Renault, Dassault, Thal` es, EZCT Cifres IFP, Renault, Thal` es (x2).

5 / 22

Stochastic continuous optimization, 2

Expensive optimization

◮ PhD Bouzarkouna ◮ PhD Loshchilov ◮ Noisy optimization

6 / 22

Stochastic continuous optimization, 3

Black-Box Optimization Benchmark Extensions to multi-objective optimization: ANR NumBBO

7 / 22

Stochastic optimization

Divide and Evolve: scaling up Planning

◮ Evolutionary computation + local planner

ICAPS 2010; IJCAI 2013

◮ Winner temporal satisficing track, IPC 2011 ◮ Winner silver award Humies GECCO 2010 ◮ Coll. Thal`

es, On´ era, Cril;

◮ PhD Cifre Bibai; ANR Descarwin

8 / 22

SIG Optimal Decision Making Under Uncertainty

From Go to energy management

◮ STREP MASH, Citines; ANR IOMCA, Explora; ADEME Post ◮ Coll. U. Tainan, Taiwan; ILAB SME Artelys; platform Metis.

9 / 22

Optimal Decision Making Under Uncertainty, 2

Some extensions

◮ Continuous spaces (PhD Couetoux) ◮ Double progressive widening (PhD Couetoux) ◮ Multi-objective MCTS (PhD Wang) ◮ MARAB: risk-aware MAB (PhD Galichet)

Applications to Machine Learning

◮ Active Learning (PhD Rolet; Digiteo 2008-2010; coll. CEA) ◮ Feature Selection (PhD Gaudel) ◮ Coll. Orange

Application to Computer Science

◮ Optimization of DFT in Spiral, coll. CMU (PhD Rimmel) ◮ Cooperation control in parallel SAT Solving, coll. Microsoft

10 / 22

Optimal Decision Making Under Uncertainty, 3

Game theory Partially observable games are undecidable even in the case of finite state spaces and deterministic transitions. Results on other games

◮ MineSweeper ◮ Havannah ◮ Urban Rivals ◮ Coll. U. Tainan; USVQ; GaLaC.

World Award: Chess Base 2009

11 / 22

Optimal Decision Making Under Uncertainty, 4

MineSweeper and MCTS

◮ All locations have same probability of

death 1/3

◮ Are then all moves equivalent ?

12 / 22

Optimal Decision Making Under Uncertainty, 4

MineSweeper and MCTS

◮ All locations have same probability of

death 1/3

◮ Are then all moves equivalent ?

NO !

12 / 22

Optimal Decision Making Under Uncertainty, 4

MineSweeper and MCTS

◮ All locations have same probability of

death 1/3

◮ Are then all moves equivalent ?

NO !

◮ Top, Bottom: Win with probability 2/3

12 / 22

Optimal Decision Making Under Uncertainty, 4

MineSweeper and MCTS

◮ All locations have same probability of

death 1/3

◮ Are then all moves equivalent ?

NO !

◮ Top, Bottom: Win with probability 2/3 ◮ MYOPIC approaches LOSE.

12 / 22

Optimal Decision Making Under Uncertainty, 5

Multi-armed bandits and Constraint Programming

◮ Coll. Microsoft-INRIA, KTH. ◮ Invited tutorial CP 2012 ◮ Dagstuhl 2014, Constraints, Optimization and Data

(co-organization with L. de Raedt, B. O’Sullivan, P. Van Hentenryck). Zoom

◮ Bandit-based Search for Constraint Programming

CP 13

13 / 22

Large scale modelling (Big data)

Context: EGEE Enabling Grids for e-Science in Europe: 100,000 + CPU; 5Pb storage; 300,000 jobs/day Data acquisition

◮ Grid Observatory portal ◮ Coll. LAL, Imperial College ◮ EGI, CNRS, INRIA, Digiteo,

UPS

14 / 22

Large scale modelling, 2

Autonomic Computing

◮ Job allocation & quality of service (reinforcement learning;

PhD Perez)

◮ Job monitoring (data streaming; PhD Zhang)

TKDE 2013

◮ Fault detection (coll. filtering; PhD Feng)

X.Z: Outstanding Award from National China Research Council for Abroad Students

Zoom on Data Streaming with Affinity Propagation

N subsets exemplars exemplars WEIGHTED AFFINITY PROPAGATION AFFINITY PROPAGATION

n2 → n1+ǫ scale invariance

PhysRev 10

15 / 22

Future

Scientific goals

◮ Multi-scale optimization under uncertainty ◮ Representation design, information theory and priors ◮ Tackling the underspecified (Human-Machine-Loop)

Initiative

◮ Data science institute proposal @ UPSay, B. K´

egl co-PI

16 / 22

Multi-scale optimization under uncertainty

Challenges

◮ Stochastic uncertainties (price, demand, weather) ◮ Large-scale problems & non-linear effects ◮ High dimension ◮ Multi-scale time horizons

POST (ADEME)

17 / 22

Representation design, information theory, priors

Deep learning

◮ Tiling the space vs learning features

Bengio, 2012

Directions

◮ Best latent marginal (PhD Arnold)

arg max qI E[log

• h

P(x|h)qD(h)] with q(h) =

• ˜

x

q(h|˜ x)PD(˜ x)

◮ Enforcing priors (PhD Isaac, coll. CEA LIST)

18 / 22

Tackling the underspecified

Reconsidering optimal decision and design

◮ Reinforcement learning: rewards are given ◮ Inverse RL: learning rewards from expert demonstrations

Directions

◮ Preference-based reinforcement learning (PhD Akrour, IP

SYMBRION)

◮ Open-ended evolution (PhD Montanier, IP SYMBRION)

19 / 22

UPSay & Data Science Institute

The 4th paradigm: Data-driven science

◮ Initially push by industry and e.g. physics and biology ◮ International initiatives: Data to Knowledge to Action; NYU;

Berkeley; Amsterdam; ... UPSay

◮ Labex DigiCosme, axis DataSense ◮ Master ML-Information & Contents

with LTCI, LIX, Evry, ECP, ENSTA

20 / 22

Visibility

International Committees

◮ ACM SIGEVO Exec. ◮ PPSN Steering ◮ PASCAL -I, -II Steering ◮ EGEE III Steering

Organisation, coordination

◮ Dagstuhl Theory of EC 08, 10 ◮ ThRaSH Wshp (2009-now) ◮ BBOB Wshp (2009-10-12-13) ◮ PASCAL Wshops and Challenges ◮ Grids Meet AC 09 ◮ EvoDeRob, IROS’09 ◮ LION 2012 ◮ Digiteo, DigiCosme, UPSay senate

Editorship

◮ Editor in Chief

Evolutionary Computation (2002-2009), Editorial Board Member GPEM, ASOC, TCS-C, EC

◮ Editorial Board Member

MLJ, GPEM, KAIS Program Committees

◮ Co-chair ECML/PKDD

2010

◮ Major ML conf (NIPS,

ICML, ECML, PKDD, IEEE-ICDM)

◮ Major EC conf. (PPSN,

ACM GECCO, IEEE CEC, Evo*) Yann Ollivier : Bronze medal CNRS

21 / 22

Apprentissage & Optimisation

MASH (Frp) Grid Observatory Peugeot EvoTest (Strep) GENNETEC (Strep) TRAVESTI Simplified Models IOMCA Innov’Nation EGEE III (IP)

APPLICATIONS

• Unsup. Brain

DigiBrain Modyrum

MACHINE LEARNING

POST Adapt (Microsoft−INRIA) SYMBRION (IP) SyDiNMaLaS CSDL DESCARWIN OMD1−2 IFP LOGIMA SIMINOLE PASCAL1 −2 (NoE) ASAP DEMAIN DigiBrain Citines (Strep) TIMCO (Bull) Orange Legend: Europe ANR Region

OPTIMISATION THEORY

Industry NumBBO

22 / 22