Section: Overall Objectives

Highlights of the Year

• Energy management is becoming one of the main focuses, and the most important applicative focus of TAO UCT-SIG. Underlying the various debates, ranging from climate change to nuclear power and integration of renewable energies in the grid (including transportation and storage), is a burning need for scenario simulation, evaluation and optimization. The scientific challenges concern the handling of continuous and discrete uncertainties (e.g. ecological impacts or emergence of future technologies) with a long term horizon. Our commitment is to provide principled studies of various investment scenarios in economical and ecological terms, including a rigorous handling of uncertainties. Specifically,

  • We actively worked to develop collaborations between European and Taiwanese experts of energy management (organization of a forum in Taiwan, http://top.twman.org/2012frtw , of meetings between French companies and Taiwanese academic visitors in Limoges and Paris http://www.lri.fr/ teytaud/france2012.html ).

  • We developed a Ilab (collaboration between Inria Saclay-IDF / Artelys) on energy, involving our common participation to the European project Citines http://www.citines.com , aimed at optimal energy management at the scale of a city or an industrial area. We also successfully applied for an ADEME project named POST, aimed at the long term (2050) optimization of the power grid in Europe and North Africa and raising hard stochastic stock management issues. Another critical issue concerns the representation of strategies enabling to combine the good long term properties of direct policy search, and the efficiency of combinatorial optimization tools for structured problems.

    Additionally, a collaboration with Inria-Chile is under discussion. We are also working on creating a company in Taiwan, working with tools from the French industry.

  • We also participated in several energy-related European meetings, including companies (section 8.5.1 ).

• Games remain a key and cool showcase to demonstrate the efficiency of our algorithms:

  Our meta-learning approach in Monte-Carlo Tree Search (MCTS) was illustrated by playing 12 games against professional players in even conditions in 7x7; it won 7 games (6/6 win with the easy side and 1/6 win with the difficult side). We achieved the best performances so far on small board minesweeper, demonstrating the efficiency of MCTS on one-player stochastic games. In collaboration with Olivier Buffet (Loria), we scaled up previous implementations to large boards, demonstrating the efficiency of Monte-Carlo Tree Search as a tool for improving existing heuristics. For illustrating the pedagogical properties of simulation-based approaches, we developed tools for generating nice test cases in games and automatically checking the opponent level.

  Besides, we realized experimental biological measurements (neuro-imagery, skin conductivity) on amateur and professional players, for further comparison and analysis.

• One of the main fundamental milestones on the TAO research agenda has been achieved by the OPT-SIG, bridging the gap between practice and theory in stochastic optimization through information-geometric optimization (IGO). IGO is devised as a canonical way to turn any smooth parametric family of probability distributions on an arbitrary, discrete or continuous search space $X$ into a continuous-time black-box optimization method on $X$. Rooted on the Fisher metric, IGO shows invariance properties under various parameterizations of the distribution family [71] , [19] , [20] . IGO covers the state-of-art CMA-ES (invariant w.r.t. monotonous transformations of the objective function and linear transformations of the coordinate space) as a special case where the probability distribution is Gaussian.

This paper got the excellent paper award (international track) at TAAI conference (given to 3/55 papers).

Best Paper Award :

[36] Strategic Choices: Small Budgets and Simple Regret in TAAI.

C.-W. Chou, P.-C. Chou, C.-S. Lee, D. Lupien Saint-Pierre, O. Teytaud, M.-H. Wang, L.-W. Wu, S.-J. Yen.