Section: Partnerships and Cooperations

International Initiatives

INDEMA

• Title: Intelligent Decision Making Mechanisms with Hidden Information, and Application to Electricity Generation

• Inria principal investigator: Olivier Teytaud

• International Partner: National University of Tainan (Taiwan)

• Duration: 2012 - 2014

• See also: http://www.lri.fr/~teytaud/taiwanday2.html

• The objectives of the project are threefolds:

  • Objective 1: Designing consistent iterative realistic algorithms for partially observable 1-player or 2-player games. We mean:

    d $•$ consistent algorithms, in the sense that they are mathematically, provably, optimal asymptotically in the computation time.

    $•$ iterative algorithms in the sense that when you give more time to the algorithm, it should be better; and with little time, it should do its best for replying something acceptable. This is also termed an anytime algorithm. Most algorithm which survive decades are iterative.

    $•$ realistic algorithms; we mean that one can easily design a consistent iterative algorithm that will never work in practice in a real-world setting; so, additionally, we want an algorithm which looks reasonnable and we refer to the second objective for the assesment of this property.

  • Objective 2: Impressive visible applications, e.g., applications in games or puzzles, because such games are very clear assessment tools. Possibilities include Minesweeper (on which we believe that much progress is still possible), Chinese Dark Chess, Kriegspiel, Phantom-Go, card games. Such nice results are critical for advertising and assessing our research.

  • Objective 3: Big industrial applications. Having both mathematics and visible realizations in games and industrial applications might be considered as too much; yet, we have chosen to request the maximum possible funding and to include many people in the travelling; also, the persons in the project are all people working in related subjects, with various terminologies, and we already have concrete applications in mind, just far enough from our past activities for being new (we want to tackle in a principled manner partial observability which was somehow ignored in many past works) and close enough for strongly reducing the ?warm up? time. In the fully observable case, we worked successfully for these three objectives and want to do the same in the partially observable case. More precisely, when working on real applications in the ?eld of energy generation, we have seen that many problems are simpli?ed so that they boil down to fully observable problems, but that this is a bad application; and our solvers must include some tricks for the partial observability. This is the main motivation for this project; we assume that mathematical analysis can be done on this (objective 1); that it will provide big results in games (objective 2) where many main programs are based on non-consistent algorithms. We believe that requirements above (objective 1) and visible realizations will facilitate the migration to real-world application; also we point out that previous research projects involving us facilitated contacts with industry, in particular in the ?eld of energy generation, which is a key point for this third objective. A roadmap for objective 3 is as follows:

    $•$ Check on simple versions of energy production problems whether the fully observable approximation is ok. We guess that in many cases it is not ok, and we want to clearly state to which extent (by how many percents) we loose in terms of loss function.

    $•$ Experiment our algorithms on real industrial problems. We will work both on Taiwan-centered and on Europe-Centered electricity generation problems in order to widen the scope of the analysis and so that both partners can be helpful in terms of applications in their own countries.

Some continuously updated and more detailed descriptions of several works in progress can be found at http://www.lri.fr/~teytaud/indema.html .

• On-going collaboration with Christian Schulte (KTH, Stockholm), one of the main developers of the GECODE Constraint Programming platform (see Section 6.2 ).

• Shinshu University, Faculty of Engineering, project Global Research on the Framework of Evolutionary Solution Search to Accelerate Innovation, from the "Strategic Young Researcher Overseas Visits Program for Accelerating Brain Circulation" program, in which TAO and DOLPHIN (Inria Lille) are partner labs and will host Japanese students in the forthcoming 4 years.