Designing and Implementing the basic data structure, genetic operators and generation of the initial populations in a learning system combining ILP and GA

Student:Ralf Haselmann
Title:Designing and Implementing the basic data structure, genetic operators and generation of the initial populations in a learning system combining ILP and GA
Type:diploma thesis
Advisors:Kókai, G.; Schneider, H.
State:submitted on November 8, 2000
Prerequisits:

ystems that induce first-order logic programs have drawn considerable interest recently within the artificial intelligence community. Inductive logic programming (ILP), for example, has very impressive applications in knowledge discovery in databases. Genetic programming (GP), a promising alternative that builds on genetic algorithm search strategies, demonstrates equally impressive results across a wide range of uses.
Both these strategies, however, have serious limitations. Despite its strong theoretical foundation from logic programming and computational learning theory, ILP does not handle concept learning well, nor can it achieve other learning paradigms such as reinforcement learning and strategy learning. GP has a much weaker theoretical foundation, as well as a laundry list of practical shortcomings.

Genetic logic programming (GLP) was apparently first suggested from Whigham as an alternative approach to Inductive Logic Programming. Lappoon et al. also implemented a version of GLP for inducing Prolog programs but his xperiments shows that to compare the performance of ILP, GP and GLP on four easy list examples the performance of GLP was the worst; it did not produce a completely correct program in any of the trials. Wong and Kwong developed Programming Structure, that also integrates these two better known approaches. To demonstrate the viability of their GLPS approach, they have tested a preliminary implementation on a battery of learning tasks: Winston's arch problem, the modified Quinlan network reachability problem, the factorial problem, and the chess endgame problem

Topic:

Developing and implementing the data structure which builds the joint representation form of the two different learning methods. On the one side it gives the form of the individuals using genetic algorithm and is generated from Prolog clauses On the other side from this representation new Prolog programs can be easily generated to examine the examples in the learning task.
Elaborate genetic operators which can be efficient applied on the data structure described above. They have warrented specialization and generatlization, to respect new background knowledge, to produce new alternative solution and so on.
Implement the generation of initial population based on the background knowledge.

watermark seal