TAO lives on the boundaries of Machine Learning and Evolutionary Computation.
For history: once upon a time there was an Evolution and Learning Group in Ecole Polytechnique (1986-2002), at the boundaries of the Center for Applied Maths and Mechanics of Solids Labs; and these guys were investigating Artificial Intelligence techniques to solve ill-posed problems, which could not be handled otherwise anyway... These guys ultimately founded TAO, bringing in their expertise in Numerical Engineering.
What is Machine Learning ?
Some answers:- An alternative to standard programmation: given examples of solved problems, the machine learns how to solve (new) problems. Very useful when the specifications of the program are unknown anyway (what the robot should do on Mars; how to play the game of Go; etc)
- The only way to deal with the overwhelming waves of data; radars, satellites, sensors,... generate tons of data, only a machine can read this gigantic and boring information in order to know what's going on (any fire in the forest ? any credit card stolen ?)
- A kind data assistant: scientists and engineers are designing experiments and gathering results; what about finding the laws (aka models) of the data ? what about suggesting hypotheses ? what about suggesting new experiments in order to confirm or reject these hypotheses ?
What is Evolutionary Computation ?
All is about optimizing an objective function. Two situations: if the objective function is nice, just follow the steepest descent direction (or the gradient) until reaching the optimum spot. Otherwise... EC is one good answer.- Construct a population (a set) of candidate solutions; usually they are not great; the point is that they should be sufficiently diverse;
- Let these solutions play together for a long time, favouring good ones, of course, and letting solutions randomly generate new solutions; how is the point;
- The game is that fittest individuals survive and reproduce: after some hundred thousand years (doesn't take that long in the computer), well-bred solutions appear.
TAO: When, Who, What, ...
TAO was born in 2003, inheriting from the ML group in Université Paris-Sud, and part of the EC group in INRIA. Why mixing these ? Because they can actually help each other:- ML is an optimization problem, and we'd better consider it an ill-posed problem.
- EC, for the sake of efficiency, should better be able to learn about the landscape it explores...
For history: once upon a time there was an Evolution and Learning Group in Ecole Polytechnique (1986-2002), at the boundaries of the Center for Applied Maths and Mechanics of Solids Labs; and these guys were investigating Artificial Intelligence techniques to solve ill-posed problems, which could not be handled otherwise anyway... These guys ultimately founded TAO, bringing in their expertise in Numerical Engineering.