| [home] - [up] |
Abstract: The extraction of logic from data, as part of data mining approaches, is becoming increasingly important. We describe a method that reliably and without tuning or parameter selection produces the desired logic functions plus probability distributions of their accuracy.
When learned logic functions replace normative logic formulations, certain problems that for normative formulations reside at the second level of the polynomial hierarchy become the logic minimization problem MINSAT. This fact could be viewed as an explanation for the curious fact that problems at the second level of the polynomial hierarchy are very hard for computers, yet are solved by the human brain with ease. Indeed, if one postulates that the human brains learns logic from data, then problems that seemingly reside at the second level of the polynomial hierarchy become MINSAT cases, which the brain is quite good at.
We discuss some applications in intelligent systems such as medical diagnosis, natural language processing, and traffic control.
Colloquium - 16:00
Abstract: Mixed Integer Programs (MIPs) are commonly solved with branch and bound algorithms based on linear programming. The success and the speed of the algorithm strongly depends on the strategy used to select the branching variables.
Today's state-of-the-art strategy is called "pseudocost branching" and uses information of previous branchings to determine the current branching.
We propose a modification of "pseudocost branching" which we call "history branching". This strategy has been implemented in SIP, a state-of-the-art MIP solver. We give computational results that show the superiority of the new strategy.