Qualitative mechanical problem-solving by artificial agents using hybrid ai

Abstract

Qualitative mechanical problem-solving (QMPS) is central to human-level intelligence. Humans use their capacity for such problem-solving in tasks as routine as opening the tap to drink water as well as for more sophisticated tasks in demanding jobs that pay well in to- day’s economy (e.g., in emergency medicine, plumbing, and the use of hydraulic machinery). Unfortunately, AIs and robots of today lack the capacity in question. This thesis investigates how qualitative mechanical problems can be solved by artificial agents and thereby takes a step toward addressing this deficiency. While statistical and connectionist ML systems have of late achieved exemplary success in producing behavior in artificial agents that resembles aspects of human intelligence, such agents are nonetheless seriously inadequate. Some inadequacies include: requiring huge datasets (often biased), which overestimate model accuracy; lack of justifications in the form of arguments/proofs behind actions performed; brittleness and unreliability that make practical-world use problematic; etc. This state-of-affairs means, e.g., that even state-of- the-art neural network ML models like GPT-3 and other powerful transformers perform poorly on reasoning problems in general and on arbitrary qualitative mechanical problems in particular. Another underlying reason for this gap is the paucity of proper datasets: there is hardly any dataset for learning to reason about qualitative mechanical problems.Under the rubric of Psychometric AI (PAI) [10], I focus on a special class of qual- itative mechanical problems: test problems from the Bennett Mechanical Comprehension Tests (BMCT-I and BMCT-II). The first part (II) of this doctoral dissertation introduces and implements a purely logicist approach for QMPS by artificial agents that is based on formalized general principles for the domain and automated reasoning. I also demonstrate its use by a robot for solving a physicalized test item in this important class of qualitative mechanical problems in the real world. In the second part (III), I devise and implement a hybrid approach to QMPS, one that draws from both statistical learning and logicist AI. My proposed methodology, based on formal reasoning, but enhanced by the transformer model technology, aims to enable artificial agents to answer Bennett and Bennett-like questions more efficiently and provide cogent justifications for their answers. Such agents are a stepping-stone on the long road toward a time when their descendants have the capacity to solve QMPS challenges in the real world at the level of human intelligence, and perhaps beyond.