The field of genetic programming (GP) is looking for hard test problems, and it has been proposed that the Tartarus problem - in which an agent moves blocks from the interior of a grid to the walls - is a suitable candidate. It is argued that, as GP solutions typically fail to find optimal solutions for this problem, it is a difficult problem.
This talk will demonstrate that, while the description of Tartarus being difficult is partially true, the main reason for GP difficulty is due to its parse tree representation (and related operators) being unable to effectively search the solution space. Once that is done, the talk will revisit the (now abandoned) use of finite state automata representations for the Tartarus problem, and demonstrate that they can be quite effective at searching the solution space.
Last modified: Tuesday, 14-Aug-2012 09:49:23 NZST
This page is maintained by the seminar list administrator.