Massive Sets of Heuristics for Machine Learning

This project aims at developing new machine learning methods relying on very large number of hand-designed heuristics, together with statistical tools to facilitate the design of these heuristics in an open and collaborative framework. We define an heuristic to be any algorithm processing raw inputs to produce values relevant to the problem at hand. This purposely very general definition encompasses techniques spanning from simple signal processing to symbolic modeling or locally trained predictors. Since we assume high performance can only be achieved by combining hundreds of such heuristics, we propose to develop them collaboratively, in a way similar to the successful development process of open-source software or collaborative encyclopedia. We will assess the performance of that strategy on the control of an avatar in a realistic 3D simulator and on the control of a real robotic arm, and we aim at creating a generic software platform usable on alternative applications. Hence, the key aspects of this proposal are to: - develop novel statistical techniques for prediction and goal-planning with a very large heterogeneous set of features, - develop statistical tools such as similarity measures in the space of features to help the design of very large sets of heuristics by many contributors, - assess the efficiency of this approach on a series of complex tasks in a realistic simulated 3D environment and with a real robot arm. The five partners of the consortium are from the fields of applied and theoretical statistical learning, reinforcement learning, artificial vision and robotics.
Machine Learning
Idiap Research Institute
Centre national de la recherche scientifique, Czech Technical University in Prague, Institut de Recherche en Informatique et en Automatique, Universitaet Potsdam
Seventh Framework Programme
Jan 01, 2010
Jun 30, 2013