Section:
Partnerships and Cooperations2>
International Initiatives3>
Inria Associate Teams4>
INDEMA5>
Title: Intelligent Decision Making Mechanisms with Hidden Information, and Application to Electricity Generation

International Partner (Institution  Laboratory  Researcher):

See also: http://www.lri.fr/~teytaud/indema.html
The objectives of the project are threefolds:
Objective 1: Designing consistent iterative realistic algorithms for partially observ
able 1player or 2player games. We mean:
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 al
gorithm that will never work in practice in a realworld setting; so, additionally, we
want an algorithm which looks reasonnable and we refer to the second objective for
the assesment of this property
We consider our work on MineSweeper[31] and combining Mcts and Dps [38] as realizations of this principle; we also apply this principle for real applications in the related Citines project .
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,
PhantomGo, card games. Such nice results are critical for advertising and assessing our
research. Since the beginning of the project, we had results on MineSweeper, Urban Rivals.
Objective 3: Big industrial applications. Having both mathematics and visible real
izations 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 ob
servability 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 field of energy generation, we have seen
that many problems are simplified 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 (objec
tive 2) where many main programs are based on nonconsistent algorithms. We believe
that requirements above (objective 1) and visible realizations will facilitate the migration
to realworld application; also we point out that previous research projects involving us
facilitated contacts with industry, in particular in the field 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
Taiwancentered and on EuropeCentered 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.
We have made papers related to energy management, including papers in very applied conferences.
We are in the process of creating a company in Taiwan, hopefully during the 2nd semester of 2013.
One student (Adrien Couëtoux) has spent 6 months there, another student has spent 5 months; Adrien just
starts a second 6 months stay there.
Inria International Partners4>
Microsoft Research Cambridge5>
Inria Associate Teams4>
INDEMA5>
Title: Intelligent Decision Making Mechanisms with Hidden Information, and Application to Electricity Generation

International Partner (Institution  Laboratory  Researcher):

See also: http://www.lri.fr/~teytaud/indema.html
The objectives of the project are threefolds:
Objective 1: Designing consistent iterative realistic algorithms for partially observ
able 1player or 2player games. We mean:
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 al
gorithm that will never work in practice in a realworld setting; so, additionally, we
want an algorithm which looks reasonnable and we refer to the second objective for
the assesment of this property
We consider our work on MineSweeper[31] and combining Mcts and Dps [38] as realizations of this principle; we also apply this principle for real applications in the related Citines project .
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,
PhantomGo, card games. Such nice results are critical for advertising and assessing our
research. Since the beginning of the project, we had results on MineSweeper, Urban Rivals.
Objective 3: Big industrial applications. Having both mathematics and visible real
izations 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 ob
servability 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 field of energy generation, we have seen
that many problems are simplified 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 (objec
tive 2) where many main programs are based on nonconsistent algorithms. We believe
that requirements above (objective 1) and visible realizations will facilitate the migration
to realworld application; also we point out that previous research projects involving us
facilitated contacts with industry, in particular in the field 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
Taiwancentered and on EuropeCentered 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.
We have made papers related to energy management, including papers in very applied conferences.
We are in the process of creating a company in Taiwan, hopefully during the 2nd semester of 2013.
One student (Adrien Couëtoux) has spent 6 months there, another student has spent 5 months; Adrien just
starts a second 6 months stay there.
Inria International Partners4>
Microsoft Research Cambridge5>
Title: Intelligent Decision Making Mechanisms with Hidden Information, and Application to Electricity Generation
International Partner (Institution  Laboratory  Researcher):
See also: http://www.lri.fr/~teytaud/indema.html
The objectives of the project are threefolds:
Objective 1: Designing consistent iterative realistic algorithms for partially observ able 1player or 2player games. We mean:
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 al gorithm that will never work in practice in a realworld setting; so, additionally, we want an algorithm which looks reasonnable and we refer to the second objective for the assesment of this property
We consider our work on MineSweeper[31] and combining Mcts and Dps [38] as realizations of this principle; we also apply this principle for real applications in the related Citines project .
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, PhantomGo, card games. Such nice results are critical for advertising and assessing our research. Since the beginning of the project, we had results on MineSweeper, Urban Rivals.
Objective 3: Big industrial applications. Having both mathematics and visible real izations 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 ob servability 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 field of energy generation, we have seen that many problems are simplified 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 (objec tive 2) where many main programs are based on nonconsistent algorithms. We believe that requirements above (objective 1) and visible realizations will facilitate the migration to realworld application; also we point out that previous research projects involving us facilitated contacts with industry, in particular in the field 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 Taiwancentered and on EuropeCentered 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.
We have made papers related to energy management, including papers in very applied conferences. We are in the process of creating a company in Taiwan, hopefully during the 2nd semester of 2013. One student (Adrien Couëtoux) has spent 6 months there, another student has spent 5 months; Adrien just starts a second 6 months stay there.
Inria International Partners4>
Microsoft Research Cambridge5>
Within the MicrosoftInria Joint Lab, the collaboration with Youssef Hamadi (Microsoft Research Cambridge), through the Adapt project, has been pursued, in spite of the departure of the 2 PhD students Alvaro Fialho and Alejandrao Arbelaez. Nadjib Lazaar and Manuel Loth have been hired as postdoc, and a new collaboration with Christian Shulte (KTH Stockholm) based on the use of Bandit algorithm within GECODE has recently given its first results [52] (see Section 3.2 ).