A Cognitive Model of Learning by Doing

In this paper an approach to learning cognitive skills from problem solving experience is presented – addressing some phenomena well known from human learning but seldom covered together in machine learning. The core of our approach is the acquisition of recursive program schemes (RPSs) by generalization-to-n over plans, using an inductive program synthesis technique. RPSs represent domain specific control knowledge, that is problem solving strategies for classes of problems. Because RPSs are abstract schemes, representing the sub-goal structure of a domain, they are suitable for analogical problem solving and learning.