Automatic Ordering of Subgoals - A Machine Learning Approach

This paper describes a learning system, LASSY1, which explores domains represented by Prolog databases, and use its acquired knowledge to increase the efficiency of a Prolog interpreter by reordering subgoals. The system creates a model of the tasks it faces and uses the model to generate informative training tasks. While performing the training tasks the system updates its inductive knowledge base which includes statistics about number of solutions and costs of various subgoals calling patterns. The collected averages are used by a subgoal ordering procedure to estimate the costs of subgoals sequences during its search for a good ordering. The paper contains a detailed analysis of the cost computation and a detailed account of the ordering procedure. Experiments done with LASSY show an improvement of performance by a factor of 10.