Research Seminar - October 24, 2002

Noise Tolerant Discovery of General Scientific Laws with Space Projection

Abstract:

How to find relevant functions efficiently from a set of instances is a very well recognized difficulty in machine learning. To achieve this goal, this seminar introduces a new system called space projection system. A set of co-operative algorithms are involved, by which functions of n-dimensional space can be discovered from given data. These algorithms include the projection algorithm, the back projection algorithm, and the basic discovery algorithm. The projection algorithm discovers all the projection functions f(x_i, x_j) = 0 in all the possible 2-dimensional spaces W_ij, (i ¹ j, 1 £ i, j £n) by applying the basic discovery algorithm. The basic discovery algorithm chooses a series of heuristic rules to find the possible nonlinear form. Successively back projection algorithms are applied to form the available functions from lower dimensions to higher dimensional spaces (from W₂ to W₃ to W₄ and up to W_n) until the goal function is found in space W_n.