Probabilistic Theory Revision

Probabilistic Theory Revision is a method of revising a domain theory to be consistent with a set of exemplars based on a proabilistic framework. Before I arrived, PTR0, a proabilistic theory revision algorithm for propositional theories, had been written and even proven to terminate and converge. A formal presentation of PTR0 can be found in the paper "Bias-Driven Revision of Logical Domain Theories" by Moshe Koppel, Ronen Feldman, and Alberto Segre.

Obviously, the next step would be to try to extend this algorithm to first order theories. Before I arrived, a reasonable start had been made in extending PTR0 to relational theories. However, there was much still to be done . . .

During the summer of 1996, I . . .

• Made a change to the relevance test.
• Invented a new inductive learning algorithm for relational theories that takes advantage of the domain theory graph structure that is used in PTR.
• Added another revision operator for a special case that my inductive learning algorithm could not correctly handle.
• Added yet another revision operator to allow for predicate invention.

A reasonable discussion of PTR0 and the necessary extensions (including mine) towards relational theories can be found in the paper Extending Probabilistic Theory Revision to First Order Theories.

The following is yet to be done in extending PTR to first order theories:

• Correctly handling rules that contain variables in the body that do not appear in the head such as grandparent(X,Y) <- parent(X,Z) ^ parent(Z,Y).
• Extending PTR to handle recursion.
• Proving termination and convergence.