A Synergistic Approach for Deep Learning and Knowledge Engineered Solutions

Machine Learning techniques offer innovative solutions to many problems from image recognition to intractably difficult games (i.e., Go). However, researchers and developers often employ machine learning techniques in isolation from knowledge-engineered approaches, typically considered passe. While knowledge-engineered solutions have limitations, they offer valuable insights about the problem space and solution pathways. Herein, we explore a hybrid approach that synergistically leverages both knowledge engineering and machine learning. The mismatch between representational layers of machine-learned systems and explicit, symbolic knowledge in engineered solutions make a hybrid approach appear infeasible. We describe a system of complementary techniques either increase system performance or decrease system implementation cost without reconceptualizing seeming incompatible strategies. This approach applies to problems wherein knowledge- engineered solution can simplify learned information, while a machine learning method can expand problem space boundaries for knowledge-engineered solutions. We illustrate this hybrid approach in Sudoku, enabling empirical assessment of engineering vs. machine learning trade-offs.