Selected Publications (last updated nov. 2013)

• How to define an "intrinsic motivation" for autonomous agents, without any ground truth (e.g. when there is no robot simulator), and how to tackle underspecified problems ?
• How to tackle meta-learning, i.e. algorithm/heuristic selection in an algorithm portfolio, and how to tune hyper-parameters depending on the problem instance at hand ?
• How to handle some sequential decision making problems in machine learning and optimmization (feature selection, active learning, surrogate learning) ?

Rewards for autonomous agents

1. APRIL: Active Preference-learning based Reinforcement Learning
   Riad Akrour; Marc Schoenauer; Michele Sebag
   ECML PKDD 2012, Springer Verlag LNCS 7524, pp. 116-131.
   In reinforcement learning, the expert might define a reward function; or demonstrate the target behaviors (inverse reinforcement learning); or give preference feedback on the behaviors demonstrated by the agent. Active learning is used to minimize the requested preference queries.
2. Sustainable cooperative coevolution with a multi-armed bandit
   Francois-Michel De Rainville, Michele Sebag, Christian Gagné, Marc Schoenauer, Denis Laurendeau.
   GECCO 2013: 1517-1524
   When two populations co-evolve, they should have commensurate computational budgets.
3. Open-Ended Evolutionary Robotics: An Information Theoretic Approach
   Pierre Delarboulas, Marc Schoenauer, Michele Sebag.
   In Parallel Problem Solving from Nature 2010 Springer Verlag LNCS, p. 334-343
   The robot computes and optimizes a criterion on-board, without any ground truth: the quantity of information in the robotic log.

Algorithm/heuristic selection and hyper-parameter tuning

1. Collaborative hyperparameter tuning
   Remi Bardenet; Mathias Brendel; Balazs Kegl; Michele Sebag
   Int. Conf. on Machine Learning, JMLR Workshop and Conference Proceedings, 28, pp. 199-207
   Rank-based learning is used to learn the performance as a function of the hyper-parameter values.

2. Bandit-based Search for Constraint Programming
   Manuel Loth; Michele Sebag; Youssef Hamadi; Marc Schoenauer
   Int. Conf. on Principles and Practice of Constraint Programming, Springer Verlag LNCS 8124, pp. 464-480
   A multi-armed bandit is used to select the variable values during the CP search.

3. Extreme Value Based Adaptive Operator Selection
   Alvaro Fialho, Luis Da Costa, Marc Schoenauer, and Michele Sebag.
   Parallel Problem Solving From Nature 2008, Springer Verlag, pages 175--184, 2008.
   How to adaptively adjust online the probability of variation operators ?

Sequential decision making in machine learning and optimization

1. Self-adaptive surrogate-assisted covariance matrix adaptation evolution strategy
   Ilya Loshchilov, Marc Schoenauer, Michele Sebag.
   GECCO 2012: 321-328
   The invariance properties w.r.t. monotonous transformation of the objective function and affine transformations of the solution space are preserved by tightly coupling CMA-ES, Ranking-SVM and the online optimization of Ranking-SVM hyper-parameters.
2. Feature Selection as a One-Player Game
   Romaric Gaudel, Michele Sebag.
   Int. Conf. on Machine Learning 2010 359-366 Feature selection is formalized as an (intractable) reinforcement learning problem, and Monte-Carlo tree search is used to approximate the corresponding optimal policy.
3. Boosting Active Learning to Optimality: A Tractable Monte-Carlo, Billiard-Based Algorithm
   Philippe Rolet, Michele Sebag, Olivier Teytaud.
   ECML PKDD 2009: 302-317 Active learning is formalized as an (intractable) reinforcement learning problem and Monte-Carlo tree search is used to approximate the corresponding optimal policy.