| dc.description.abstract | Humans have the ability to form strategies when faced with novel tasks, and different people often form different strategies for the same task. Most AI systems, on the other hand, do not exhibit this level of fluidity. Also, there is currently limited research on how to formally represent a space of strategies such that individual strategies can be systematically synthesized, scrutinized, and transferred. This limitation acts as a barrier to building AI systems that reason fluidly.
In this dissertation, I explored three ways in which strategies could be expressed in intelligent systems. First, I investigated how strategies could be expressed in information processing terms in intelligent systems that relied on visual imagery based representations. I built models to reason through problems on the Punched-hole Paper Folding Task, the Leiter International Intelligence Scale-Revised (Leiter-R), and the Block Design Task (BDT), all of which are non-verbal intelligence reasoning tasks.
Second, I investigated how AI systems could leverage techniques from program synthesis to generate their own strategies when faced with tasks. I created a domain specific language, called Visual Imagery Reasoning Language (VIMRL), with which I built AI systems that synthesized reasoning strategies for solving tasks on the Abstract Reasoning Corpus (an abstract reasoning benchmark for AI systems), and the Block Design Task.
Third, I contributed work towards automating of the measurement of human performance on the Block Design Task. This work led to the creation of innovative systems that produce detailed recordings of human performance on the block design task, for both in-person and web based online sessions. I additionally built tools to identify and visualize strategy variations in data collected from these systems.
In this dissertation I mainly contribute the following: (1) a demonstration of the sufficiency of imagery as a viable representation for reasoning about the BDT, Leiter-R, and Paper Folding Task; (2) VIMRL, an imagery inspired domain specific language, and its associated program synthesis system that generates strategies for certain visual reasoning tasks; (3) a VIMRL based solver for the Abstract Reasoning Corpus that tied for 4th place on the 2022 international ARCATHON competition; (4) a demonstration of how reasoning strategies for simplified forms of the BDT could be synthesized and represented in VIMRL; (5) tools for recording, visualizing, and categorizing human performance and strategy on the Block Design Task. | |
