Example-guided optimization of recursive domain theories

The authors investigate the utility of explanation-based learning in recursive domain theories and examine the cost of using macro-rules in these theories. As a first step in producing effective explanation-based generalization (EBG) algorithms, the authors present a new algorithm for performing source optimization of recursive domain theories. The algorithm, RSG (recursive-structure generalizer), uses a training example as bias and generalizes the control knowledge encoded in the example's derivation tree to produce a more efficient formulation of the original domain theory. The control knowledge involves control of both clause and binding selection. The authors demonstrate the effectiveness of the method of planning problems in situation calculus. The authors show that in most cases one must know the future problem distribution a priori to produce an optimal reformulation.