TAO position

Tao's goals, pertaining to the field of Computer Science, must be understood in the light of bounded resource reasoning. Quoting Herbert Simon (1958):
''In complex real-world situations, optimization becomes approximate optimization since the description of the real-world is radically simplified until reduced to a degree of complication that the decision maker can handle.
Satisficing seeks simplification in a somewhat different direction, retaining more of the detail of the real-world situation, but settling for a satisfactory, rather than approximate-best, decision.''

Tao's contributions in Stochastic Optimization (EC) and Machine Learning (ML) definitely follow the satisficing path - forging concepts and tools for tackling ill-posed problems directly, as opposed to forcing them in a well-posed setting. Among the main scientific questions investigated are:

  • Search Spaces. In both EC and Artificial Intelligence, one is halfway through the solution when an appropriate problem space has been defined. Accordingly, various search spaces have been investigated. For instance shape design has been tackled using representations ranging from Voronoï to developmental settings and Echo State Networks. Some of these contributed to breakthrough results in the field of architecture (coll. EZCT).

  • Scalability, pervasive to Computer Science, has been tackled along two dimensions: the size of the search space (as in MoGo, since the branching factor of the search tree in Go is about 200 as opposed to circa 45 for the game of Chess; the current versions of Mogo use 30,000 simulations per move), and the input size (mining the logs of the EGEE Grid means dealing with 5Gb data).

  • Integrated Information Cycle. Three main tasks in domains ranging from robotics to simplified models are:
    1. information gathering (active learning);
    2. model building (learning);
    3. policy design (optimization).

    While these tasks are most often considered and tackled in isolation, there is a need to consider them in an integrated fashion, which is investigated along OMD (Renault, Dassault, EADS and a few PMEs) and in collaboration with CEA.

  • Algorithm portfolio. Negative results (in ML or in EC) show the search for the ``killer'' universal algorithm to be doomed to failure. Another perspective, inspired from the Constraint Satisfaction community, is based on estimating the algorithm accuracy as a random variable depending on the order parameters of the problem instances. Along this perspective, an algorithm thus comes with its expected empirical accuracy, enabling to provide the end-user with an autonomic algorithm portfolio.

Besides, approaches based on Statistical Learning Theory have let to theoretical breakthrough in Evolutionary Computation theory, known to be lagging far behind practice.

Contributors to this page: evomarc .
Page last modified on Thursday 05 of February, 2009 23:36:04 CET by evomarc.