Qualitative Mechanical Problem-Solving by Artificial Agents:

Abstract

Qualitative mechanical problem-solving (QMPS) is central to human-level intelligence. Human agents use their capacity for such problem-solving to succeed in tasks as routine as opening the tap to drink or hanging a picture on the wall, as well as for more sophisticated tasks in demanding jobs in today’s economy (e.g., emergency medicine, plumbing, hydraulic machinery, & driving). Unfortunately, artificial agents (including specifically robots) of today lack the capacity in question. Our work takes QMPS to fall under the general, longstanding AI area of qualitative reasoning (QR), historically an intensely logic-based affair. We embrace this history, and take new, further steps to advance QMPS. The Bennett Mechanical Comprehension Tests (BMCT-I and BMCT-II) assess a human’s ability to solve QMPS problems, and are used in the real world by many employers to evaluate job candidates. Building on the work of others who have attacked BMCT under the rubric of Psychometric AI (PAI), we introduce one of our novel algorithms (A_B1) in a family (A_B) of such for QMPS as required by BMCT, illustrate via case studies, report time-based performance of A_B1, and assess our progress with an eye to future work in which our approach is extended to a sub-class of algorithms in A_B that exploit the power of argument-based nonmonotonic logic, and leverage the success of transformer models to enhance their efficiency.